fs/btrfs/tree-log.c

   1 /*
   2  * Copyright (C) 2008 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/sched.h>
  20 #include <linux/slab.h>
  21 #include <linux/list_sort.h>
  22 #include "ctree.h"
  23 #include "transaction.h"
  24 #include "disk-io.h"
  25 #include "locking.h"
  26 #include "print-tree.h"
  27 #include "backref.h"
  28 #include "compat.h"
  29 #include "tree-log.h"
  30 #include "hash.h"
  31
  32 /* magic values for the inode_only field in btrfs_log_inode:
  33  *
  34  * LOG_INODE_ALL means to log everything
  35  * LOG_INODE_EXISTS means to log just enough to recreate the inode
  36  * during log replay
  37  */
  38 #define LOG_INODE_ALL 0
  39 #define LOG_INODE_EXISTS 1
  40
  41 /*
  42  * directory trouble cases
  43  *
  44  * 1) on rename or unlink, if the inode being unlinked isn't in the fsync
  45  * log, we must force a full commit before doing an fsync of the directory
  46  * where the unlink was done.
  47  * ---> record transid of last unlink/rename per directory
  48  *
  49  * mkdir foo/some_dir
  50  * normal commit
  51  * rename foo/some_dir foo2/some_dir
  52  * mkdir foo/some_dir
  53  * fsync foo/some_dir/some_file
  54  *
  55  * The fsync above will unlink the original some_dir without recording
  56  * it in its new location (foo2).  After a crash, some_dir will be gone
  57  * unless the fsync of some_file forces a full commit
  58  *
  59  * 2) we must log any new names for any file or dir that is in the fsync
  60  * log. ---> check inode while renaming/linking.
  61  *
  62  * 2a) we must log any new names for any file or dir during rename
  63  * when the directory they are being removed from was logged.
  64  * ---> check inode and old parent dir during rename
  65  *
  66  *  2a is actually the more important variant.  With the extra logging
  67  *  a crash might unlink the old name without recreating the new one
  68  *
  69  * 3) after a crash, we must go through any directories with a link count
  70  * of zero and redo the rm -rf
  71  *
  72  * mkdir f1/foo
  73  * normal commit
  74  * rm -rf f1/foo
  75  * fsync(f1)
  76  *
  77  * The directory f1 was fully removed from the FS, but fsync was never
  78  * called on f1, only its parent dir.  After a crash the rm -rf must
  79  * be replayed.  This must be able to recurse down the entire
  80  * directory tree.  The inode link count fixup code takes care of the
  81  * ugly details.
  82  */
  83
  84 /*
  85  * stages for the tree walking.  The first
  86  * stage (0) is to only pin down the blocks we find
  87  * the second stage (1) is to make sure that all the inodes
  88  * we find in the log are created in the subvolume.
  89  *
  90  * The last stage is to deal with directories and links and extents
  91  * and all the other fun semantics
  92  */
  93 #define LOG_WALK_PIN_ONLY 0
  94 #define LOG_WALK_REPLAY_INODES 1
  95 #define LOG_WALK_REPLAY_ALL 2
  96
  97 static int btrfs_log_inode(struct btrfs_trans_handle *trans,
  98                              struct btrfs_root *root, struct inode *inode,
  99                              int inode_only);
 100 static int link_to_fixup_dir(struct btrfs_trans_handle *trans,
 101                              struct btrfs_root *root,
 102                              struct btrfs_path *path, u64 objectid);
 103 static noinline int replay_dir_deletes(struct btrfs_trans_handle *trans,
 104                                        struct btrfs_root *root,
 105                                        struct btrfs_root *log,
 106                                        struct btrfs_path *path,
 107                                        u64 dirid, int del_all);
 108
 109 /*
 110  * tree logging is a special write ahead log used to make sure that
 111  * fsyncs and O_SYNCs can happen without doing full tree commits.
 112  *
 113  * Full tree commits are expensive because they require commonly
 114  * modified blocks to be recowed, creating many dirty pages in the
 115  * extent tree an 4x-6x higher write load than ext3.
 116  *
 117  * Instead of doing a tree commit on every fsync, we use the
 118  * key ranges and transaction ids to find items for a given file or directory
 119  * that have changed in this transaction.  Those items are copied into
 120  * a special tree (one per subvolume root), that tree is written to disk
 121  * and then the fsync is considered complete.
 122  *
 123  * After a crash, items are copied out of the log-tree back into the
 124  * subvolume tree.  Any file data extents found are recorded in the extent
 125  * allocation tree, and the log-tree freed.
 126  *
 127  * The log tree is read three times, once to pin down all the extents it is
 128  * using in ram and once, once to create all the inodes logged in the tree
 129  * and once to do all the other items.
 130  */
 131
 132 /*
 133  * start a sub transaction and setup the log tree
 134  * this increments the log tree writer count to make the people
 135  * syncing the tree wait for us to finish
 136  */
 137 static int start_log_trans(struct btrfs_trans_handle *trans,
 138                            struct btrfs_root *root)
 139 {
 140         int ret;
 141         int err = 0;
 142
 143         mutex_lock(&root->log_mutex);
 144         if (root->log_root) {
 145                 if (!root->log_start_pid) {
 146                         root->log_start_pid = current->pid;
 147                         root->log_multiple_pids = false;
 148                 } else if (root->log_start_pid != current->pid) {
 149                         root->log_multiple_pids = true;
 150                 }
 151
 152                 atomic_inc(&root->log_batch);
 153                 atomic_inc(&root->log_writers);
 154                 mutex_unlock(&root->log_mutex);
 155                 return 0;
 156         }
 157         root->log_multiple_pids = false;
 158         root->log_start_pid = current->pid;
 159         mutex_lock(&root->fs_info->tree_log_mutex);
 160         if (!root->fs_info->log_root_tree) {
 161                 ret = btrfs_init_log_root_tree(trans, root->fs_info);
 162                 if (ret)
 163                         err = ret;
 164         }
 165         if (err == 0 && !root->log_root) {
 166                 ret = btrfs_add_log_tree(trans, root);
 167                 if (ret)
 168                         err = ret;
 169         }
 170         mutex_unlock(&root->fs_info->tree_log_mutex);
 171         atomic_inc(&root->log_batch);
 172         atomic_inc(&root->log_writers);
 173         mutex_unlock(&root->log_mutex);
 174         return err;
 175 }
 176
 177 /*
 178  * returns 0 if there was a log transaction running and we were able
 179  * to join, or returns -ENOENT if there were not transactions
 180  * in progress
 181  */
 182 static int join_running_log_trans(struct btrfs_root *root)
 183 {
 184         int ret = -ENOENT;
 185
 186         smp_mb();
 187         if (!root->log_root)
 188                 return -ENOENT;
 189
 190         mutex_lock(&root->log_mutex);
 191         if (root->log_root) {
 192                 ret = 0;
 193                 atomic_inc(&root->log_writers);
 194         }
 195         mutex_unlock(&root->log_mutex);
 196         return ret;
 197 }
 198
 199 /*
 200  * This either makes the current running log transaction wait
 201  * until you call btrfs_end_log_trans() or it makes any future
 202  * log transactions wait until you call btrfs_end_log_trans()
 203  */
 204 int btrfs_pin_log_trans(struct btrfs_root *root)
 205 {
 206         int ret = -ENOENT;
 207
 208         mutex_lock(&root->log_mutex);
 209         atomic_inc(&root->log_writers);
 210         mutex_unlock(&root->log_mutex);
 211         return ret;
 212 }
 213
 214 /*
 215  * indicate we're done making changes to the log tree
 216  * and wake up anyone waiting to do a sync
 217  */
 218 void btrfs_end_log_trans(struct btrfs_root *root)
 219 {
 220         if (atomic_dec_and_test(&root->log_writers)) {
 221                 smp_mb();
 222                 if (waitqueue_active(&root->log_writer_wait))
 223                         wake_up(&root->log_writer_wait);
 224         }
 225 }
 226
 227
 228 /*
 229  * the walk control struct is used to pass state down the chain when
 230  * processing the log tree.  The stage field tells us which part
 231  * of the log tree processing we are currently doing.  The others
 232  * are state fields used for that specific part
 233  */
 234 struct walk_control {
 235         /* should we free the extent on disk when done?  This is used
 236          * at transaction commit time while freeing a log tree
 237          */
 238         int free;
 239
 240         /* should we write out the extent buffer?  This is used
 241          * while flushing the log tree to disk during a sync
 242          */
 243         int write;
 244
 245         /* should we wait for the extent buffer io to finish?  Also used
 246          * while flushing the log tree to disk for a sync
 247          */
 248         int wait;
 249
 250         /* pin only walk, we record which extents on disk belong to the
 251          * log trees
 252          */
 253         int pin;
 254
 255         /* what stage of the replay code we're currently in */
 256         int stage;
 257
 258         /* the root we are currently replaying */
 259         struct btrfs_root *replay_dest;
 260
 261         /* the trans handle for the current replay */
 262         struct btrfs_trans_handle *trans;
 263
 264         /* the function that gets used to process blocks we find in the
 265          * tree.  Note the extent_buffer might not be up to date when it is
 266          * passed in, and it must be checked or read if you need the data
 267          * inside it
 268          */
 269         int (*process_func)(struct btrfs_root *log, struct extent_buffer *eb,
 270                             struct walk_control *wc, u64 gen);
 271 };
 272
 273 /*
 274  * process_func used to pin down extents, write them or wait on them
 275  */
 276 static int process_one_buffer(struct btrfs_root *log,
 277                               struct extent_buffer *eb,
 278                               struct walk_control *wc, u64 gen)
 279 {
 280         if (wc->pin)
 281                 btrfs_pin_extent_for_log_replay(log->fs_info->extent_root,
 282                                                 eb->start, eb->len);
 283
 284         if (btrfs_buffer_uptodate(eb, gen, 0)) {
 285                 if (wc->write)
 286                         btrfs_write_tree_block(eb);
 287                 if (wc->wait)
 288                         btrfs_wait_tree_block_writeback(eb);
 289         }
 290         return 0;
 291 }
 292
 293 /*
 294  * Item overwrite used by replay and tree logging.  eb, slot and key all refer
 295  * to the src data we are copying out.
 296  *
 297  * root is the tree we are copying into, and path is a scratch
 298  * path for use in this function (it should be released on entry and
 299  * will be released on exit).
 300  *
 301  * If the key is already in the destination tree the existing item is
 302  * overwritten.  If the existing item isn't big enough, it is extended.
 303  * If it is too large, it is truncated.
 304  *
 305  * If the key isn't in the destination yet, a new item is inserted.
 306  */
 307 static noinline int overwrite_item(struct btrfs_trans_handle *trans,
 308                                    struct btrfs_root *root,
 309                                    struct btrfs_path *path,
 310                                    struct extent_buffer *eb, int slot,
 311                                    struct btrfs_key *key)
 312 {
 313         int ret;
 314         u32 item_size;
 315         u64 saved_i_size = 0;
 316         int save_old_i_size = 0;
 317         unsigned long src_ptr;
 318         unsigned long dst_ptr;
 319         int overwrite_root = 0;
 320
 321         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID)
 322                 overwrite_root = 1;
 323
 324         item_size = btrfs_item_size_nr(eb, slot);
 325         src_ptr = btrfs_item_ptr_offset(eb, slot);
 326
 327         /* look for the key in the destination tree */
 328         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
 329         if (ret == 0) {
 330                 char *src_copy;
 331                 char *dst_copy;
 332                 u32 dst_size = btrfs_item_size_nr(path->nodes[0],
 333                                                   path->slots[0]);
 334                 if (dst_size != item_size)
 335                         goto insert;
 336
 337                 if (item_size == 0) {
 338                         btrfs_release_path(path);
 339                         return 0;
 340                 }
 341                 dst_copy = kmalloc(item_size, GFP_NOFS);
 342                 src_copy = kmalloc(item_size, GFP_NOFS);
 343                 if (!dst_copy || !src_copy) {
 344                         btrfs_release_path(path);
 345                         kfree(dst_copy);
 346                         kfree(src_copy);
 347                         return -ENOMEM;
 348                 }
 349
 350                 read_extent_buffer(eb, src_copy, src_ptr, item_size);
 351
 352                 dst_ptr = btrfs_item_ptr_offset(path->nodes[0], path->slots[0]);
 353                 read_extent_buffer(path->nodes[0], dst_copy, dst_ptr,
 354                                    item_size);
 355                 ret = memcmp(dst_copy, src_copy, item_size);
 356
 357                 kfree(dst_copy);
 358                 kfree(src_copy);
 359                 /*
 360                  * they have the same contents, just return, this saves
 361                  * us from cowing blocks in the destination tree and doing
 362                  * extra writes that may not have been done by a previous
 363                  * sync
 364                  */
 365                 if (ret == 0) {
 366                         btrfs_release_path(path);
 367                         return 0;
 368                 }
 369
 370         }
 371 insert:
 372         btrfs_release_path(path);
 373         /* try to insert the key into the destination tree */
 374         ret = btrfs_insert_empty_item(trans, root, path,
 375                                       key, item_size);
 376
 377         /* make sure any existing item is the correct size */
 378         if (ret == -EEXIST) {
 379                 u32 found_size;
 380                 found_size = btrfs_item_size_nr(path->nodes[0],
 381                                                 path->slots[0]);
 382                 if (found_size > item_size)
 383                         btrfs_truncate_item(trans, root, path, item_size, 1);
 384                 else if (found_size < item_size)
 385                         btrfs_extend_item(trans, root, path,
 386                                           item_size - found_size);
 387         } else if (ret) {
 388                 return ret;
 389         }
 390         dst_ptr = btrfs_item_ptr_offset(path->nodes[0],
 391                                         path->slots[0]);
 392
 393         /* don't overwrite an existing inode if the generation number
 394          * was logged as zero.  This is done when the tree logging code
 395          * is just logging an inode to make sure it exists after recovery.
 396          *
 397          * Also, don't overwrite i_size on directories during replay.
 398          * log replay inserts and removes directory items based on the
 399          * state of the tree found in the subvolume, and i_size is modified
 400          * as it goes
 401          */
 402         if (key->type == BTRFS_INODE_ITEM_KEY && ret == -EEXIST) {
 403                 struct btrfs_inode_item *src_item;
 404                 struct btrfs_inode_item *dst_item;
 405
 406                 src_item = (struct btrfs_inode_item *)src_ptr;
 407                 dst_item = (struct btrfs_inode_item *)dst_ptr;
 408
 409                 if (btrfs_inode_generation(eb, src_item) == 0)
 410                         goto no_copy;
 411
 412                 if (overwrite_root &&
 413                     S_ISDIR(btrfs_inode_mode(eb, src_item)) &&
 414                     S_ISDIR(btrfs_inode_mode(path->nodes[0], dst_item))) {
 415                         save_old_i_size = 1;
 416                         saved_i_size = btrfs_inode_size(path->nodes[0],
 417                                                         dst_item);
 418                 }
 419         }
 420
 421         copy_extent_buffer(path->nodes[0], eb, dst_ptr,
 422                            src_ptr, item_size);
 423
 424         if (save_old_i_size) {
 425                 struct btrfs_inode_item *dst_item;
 426                 dst_item = (struct btrfs_inode_item *)dst_ptr;
 427                 btrfs_set_inode_size(path->nodes[0], dst_item, saved_i_size);
 428         }
 429
 430         /* make sure the generation is filled in */
 431         if (key->type == BTRFS_INODE_ITEM_KEY) {
 432                 struct btrfs_inode_item *dst_item;
 433                 dst_item = (struct btrfs_inode_item *)dst_ptr;
 434                 if (btrfs_inode_generation(path->nodes[0], dst_item) == 0) {
 435                         btrfs_set_inode_generation(path->nodes[0], dst_item,
 436                                                    trans->transid);
 437                 }
 438         }
 439 no_copy:
 440         btrfs_mark_buffer_dirty(path->nodes[0]);
 441         btrfs_release_path(path);
 442         return 0;
 443 }
 444
 445 /*
 446  * simple helper to read an inode off the disk from a given root
 447  * This can only be called for subvolume roots and not for the log
 448  */
 449 static noinline struct inode *read_one_inode(struct btrfs_root *root,
 450                                              u64 objectid)
 451 {
 452         struct btrfs_key key;
 453         struct inode *inode;
 454
 455         key.objectid = objectid;
 456         key.type = BTRFS_INODE_ITEM_KEY;
 457         key.offset = 0;
 458         inode = btrfs_iget(root->fs_info->sb, &key, root, NULL);
 459         if (IS_ERR(inode)) {
 460                 inode = NULL;
 461         } else if (is_bad_inode(inode)) {
 462                 iput(inode);
 463                 inode = NULL;
 464         }
 465         return inode;
 466 }
 467
 468 /* replays a single extent in 'eb' at 'slot' with 'key' into the
 469  * subvolume 'root'.  path is released on entry and should be released
 470  * on exit.
 471  *
 472  * extents in the log tree have not been allocated out of the extent
 473  * tree yet.  So, this completes the allocation, taking a reference
 474  * as required if the extent already exists or creating a new extent
 475  * if it isn't in the extent allocation tree yet.
 476  *
 477  * The extent is inserted into the file, dropping any existing extents
 478  * from the file that overlap the new one.
 479  */
 480 static noinline int replay_one_extent(struct btrfs_trans_handle *trans,
 481                                       struct btrfs_root *root,
 482                                       struct btrfs_path *path,
 483                                       struct extent_buffer *eb, int slot,
 484                                       struct btrfs_key *key)
 485 {
 486         int found_type;
 487         u64 mask = root->sectorsize - 1;
 488         u64 extent_end;
 489         u64 start = key->offset;
 490         u64 saved_nbytes;
 491         struct btrfs_file_extent_item *item;
 492         struct inode *inode = NULL;
 493         unsigned long size;
 494         int ret = 0;
 495
 496         item = btrfs_item_ptr(eb, slot, struct btrfs_file_extent_item);
 497         found_type = btrfs_file_extent_type(eb, item);
 498
 499         if (found_type == BTRFS_FILE_EXTENT_REG ||
 500             found_type == BTRFS_FILE_EXTENT_PREALLOC)
 501                 extent_end = start + btrfs_file_extent_num_bytes(eb, item);
 502         else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
 503                 size = btrfs_file_extent_inline_len(eb, item);
 504                 extent_end = (start + size + mask) & ~mask;
 505         } else {
 506                 ret = 0;
 507                 goto out;
 508         }
 509
 510         inode = read_one_inode(root, key->objectid);
 511         if (!inode) {
 512                 ret = -EIO;
 513                 goto out;
 514         }
 515
 516         /*
 517          * first check to see if we already have this extent in the
 518          * file.  This must be done before the btrfs_drop_extents run
 519          * so we don't try to drop this extent.
 520          */
 521         ret = btrfs_lookup_file_extent(trans, root, path, btrfs_ino(inode),
 522                                        start, 0);
 523
 524         if (ret == 0 &&
 525             (found_type == BTRFS_FILE_EXTENT_REG ||
 526              found_type == BTRFS_FILE_EXTENT_PREALLOC)) {
 527                 struct btrfs_file_extent_item cmp1;
 528                 struct btrfs_file_extent_item cmp2;
 529                 struct btrfs_file_extent_item *existing;
 530                 struct extent_buffer *leaf;
 531
 532                 leaf = path->nodes[0];
 533                 existing = btrfs_item_ptr(leaf, path->slots[0],
 534                                           struct btrfs_file_extent_item);
 535
 536                 read_extent_buffer(eb, &cmp1, (unsigned long)item,
 537                                    sizeof(cmp1));
 538                 read_extent_buffer(leaf, &cmp2, (unsigned long)existing,
 539                                    sizeof(cmp2));
 540
 541                 /*
 542                  * we already have a pointer to this exact extent,
 543                  * we don't have to do anything
 544                  */
 545                 if (memcmp(&cmp1, &cmp2, sizeof(cmp1)) == 0) {
 546                         btrfs_release_path(path);
 547                         goto out;
 548                 }
 549         }
 550         btrfs_release_path(path);
 551
 552         saved_nbytes = inode_get_bytes(inode);
 553         /* drop any overlapping extents */
 554         ret = btrfs_drop_extents(trans, root, inode, start, extent_end, 1);
 555         BUG_ON(ret);
 556
 557         if (found_type == BTRFS_FILE_EXTENT_REG ||
 558             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
 559                 u64 offset;
 560                 unsigned long dest_offset;
 561                 struct btrfs_key ins;
 562
 563                 ret = btrfs_insert_empty_item(trans, root, path, key,
 564                                               sizeof(*item));
 565                 BUG_ON(ret);
 566                 dest_offset = btrfs_item_ptr_offset(path->nodes[0],
 567                                                     path->slots[0]);
 568                 copy_extent_buffer(path->nodes[0], eb, dest_offset,
 569                                 (unsigned long)item,  sizeof(*item));
 570
 571                 ins.objectid = btrfs_file_extent_disk_bytenr(eb, item);
 572                 ins.offset = btrfs_file_extent_disk_num_bytes(eb, item);
 573                 ins.type = BTRFS_EXTENT_ITEM_KEY;
 574                 offset = key->offset - btrfs_file_extent_offset(eb, item);
 575
 576                 if (ins.objectid > 0) {
 577                         u64 csum_start;
 578                         u64 csum_end;
 579                         LIST_HEAD(ordered_sums);
 580                         /*
 581                          * is this extent already allocated in the extent
 582                          * allocation tree?  If so, just add a reference
 583                          */
 584                         ret = btrfs_lookup_extent(root, ins.objectid,
 585                                                 ins.offset);
 586                         if (ret == 0) {
 587                                 ret = btrfs_inc_extent_ref(trans, root,
 588                                                 ins.objectid, ins.offset,
 589                                                 0, root->root_key.objectid,
 590                                                 key->objectid, offset, 0);
 591                                 BUG_ON(ret);
 592                         } else {
 593                                 /*
 594                                  * insert the extent pointer in the extent
 595                                  * allocation tree
 596                                  */
 597                                 ret = btrfs_alloc_logged_file_extent(trans,
 598                                                 root, root->root_key.objectid,
 599                                                 key->objectid, offset, &ins);
 600                                 BUG_ON(ret);
 601                         }
 602                         btrfs_release_path(path);
 603
 604                         if (btrfs_file_extent_compression(eb, item)) {
 605                                 csum_start = ins.objectid;
 606                                 csum_end = csum_start + ins.offset;
 607                         } else {
 608                                 csum_start = ins.objectid +
 609                                         btrfs_file_extent_offset(eb, item);
 610                                 csum_end = csum_start +
 611                                         btrfs_file_extent_num_bytes(eb, item);
 612                         }
 613
 614                         ret = btrfs_lookup_csums_range(root->log_root,
 615                                                 csum_start, csum_end - 1,
 616                                                 &ordered_sums, 0);
 617                         BUG_ON(ret);
 618                         while (!list_empty(&ordered_sums)) {
 619                                 struct btrfs_ordered_sum *sums;
 620                                 sums = list_entry(ordered_sums.next,
 621                                                 struct btrfs_ordered_sum,
 622                                                 list);
 623                                 ret = btrfs_csum_file_blocks(trans,
 624                                                 root->fs_info->csum_root,
 625                                                 sums);
 626                                 BUG_ON(ret);
 627                                 list_del(&sums->list);
 628                                 kfree(sums);
 629                         }
 630                 } else {
 631                         btrfs_release_path(path);
 632                 }
 633         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
 634                 /* inline extents are easy, we just overwrite them */
 635                 ret = overwrite_item(trans, root, path, eb, slot, key);
 636                 BUG_ON(ret);
 637         }
 638
 639         inode_set_bytes(inode, saved_nbytes);
 640         ret = btrfs_update_inode(trans, root, inode);
 641 out:
 642         if (inode)
 643                 iput(inode);
 644         return ret;
 645 }
 646
 647 /*
 648  * when cleaning up conflicts between the directory names in the
 649  * subvolume, directory names in the log and directory names in the
 650  * inode back references, we may have to unlink inodes from directories.
 651  *
 652  * This is a helper function to do the unlink of a specific directory
 653  * item
 654  */
 655 static noinline int drop_one_dir_item(struct btrfs_trans_handle *trans,
 656                                       struct btrfs_root *root,
 657                                       struct btrfs_path *path,
 658                                       struct inode *dir,
 659                                       struct btrfs_dir_item *di)
 660 {
 661         struct inode *inode;
 662         char *name;
 663         int name_len;
 664         struct extent_buffer *leaf;
 665         struct btrfs_key location;
 666         int ret;
 667
 668         leaf = path->nodes[0];
 669
 670         btrfs_dir_item_key_to_cpu(leaf, di, &location);
 671         name_len = btrfs_dir_name_len(leaf, di);
 672         name = kmalloc(name_len, GFP_NOFS);
 673         if (!name)
 674                 return -ENOMEM;
 675
 676         read_extent_buffer(leaf, name, (unsigned long)(di + 1), name_len);
 677         btrfs_release_path(path);
 678
 679         inode = read_one_inode(root, location.objectid);
 680         if (!inode) {
 681                 kfree(name);
 682                 return -EIO;
 683         }
 684
 685         ret = link_to_fixup_dir(trans, root, path, location.objectid);
 686         BUG_ON(ret);
 687
 688         ret = btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
 689         BUG_ON(ret);
 690         kfree(name);
 691
 692         iput(inode);
 693
 694         btrfs_run_delayed_items(trans, root);
 695         return ret;
 696 }
 697
 698 /*
 699  * helper function to see if a given name and sequence number found
 700  * in an inode back reference are already in a directory and correctly
 701  * point to this inode
 702  */
 703 static noinline int inode_in_dir(struct btrfs_root *root,
 704                                  struct btrfs_path *path,
 705                                  u64 dirid, u64 objectid, u64 index,
 706                                  const char *name, int name_len)
 707 {
 708         struct btrfs_dir_item *di;
 709         struct btrfs_key location;
 710         int match = 0;
 711
 712         di = btrfs_lookup_dir_index_item(NULL, root, path, dirid,
 713                                          index, name, name_len, 0);
 714         if (di && !IS_ERR(di)) {
 715                 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &location);
 716                 if (location.objectid != objectid)
 717                         goto out;
 718         } else
 719                 goto out;
 720         btrfs_release_path(path);
 721
 722         di = btrfs_lookup_dir_item(NULL, root, path, dirid, name, name_len, 0);
 723         if (di && !IS_ERR(di)) {
 724                 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &location);
 725                 if (location.objectid != objectid)
 726                         goto out;
 727         } else
 728                 goto out;
 729         match = 1;
 730 out:
 731         btrfs_release_path(path);
 732         return match;
 733 }
 734
 735 /*
 736  * helper function to check a log tree for a named back reference in
 737  * an inode.  This is used to decide if a back reference that is
 738  * found in the subvolume conflicts with what we find in the log.
 739  *
 740  * inode backreferences may have multiple refs in a single item,
 741  * during replay we process one reference at a time, and we don't
 742  * want to delete valid links to a file from the subvolume if that
 743  * link is also in the log.
 744  */
 745 static noinline int backref_in_log(struct btrfs_root *log,
 746                                    struct btrfs_key *key,
 747                                    u64 ref_objectid,
 748                                    char *name, int namelen)
 749 {
 750         struct btrfs_path *path;
 751         struct btrfs_inode_ref *ref;
 752         unsigned long ptr;
 753         unsigned long ptr_end;
 754         unsigned long name_ptr;
 755         int found_name_len;
 756         int item_size;
 757         int ret;
 758         int match = 0;
 759
 760         path = btrfs_alloc_path();
 761         if (!path)
 762                 return -ENOMEM;
 763
 764         ret = btrfs_search_slot(NULL, log, key, path, 0, 0);
 765         if (ret != 0)
 766                 goto out;
 767
 768         ptr = btrfs_item_ptr_offset(path->nodes[0], path->slots[0]);
 769
 770         if (key->type == BTRFS_INODE_EXTREF_KEY) {
 771                 if (btrfs_find_name_in_ext_backref(path, ref_objectid,
 772                                                    name, namelen, NULL))
 773                         match = 1;
 774
 775                 goto out;
 776         }
 777
 778         item_size = btrfs_item_size_nr(path->nodes[0], path->slots[0]);
 779         ptr_end = ptr + item_size;
 780         while (ptr < ptr_end) {
 781                 ref = (struct btrfs_inode_ref *)ptr;
 782                 found_name_len = btrfs_inode_ref_name_len(path->nodes[0], ref);
 783                 if (found_name_len == namelen) {
 784                         name_ptr = (unsigned long)(ref + 1);
 785                         ret = memcmp_extent_buffer(path->nodes[0], name,
 786                                                    name_ptr, namelen);
 787                         if (ret == 0) {
 788                                 match = 1;
 789                                 goto out;
 790                         }
 791                 }
 792                 ptr = (unsigned long)(ref + 1) + found_name_len;
 793         }
 794 out:
 795         btrfs_free_path(path);
 796         return match;
 797 }
 798
 799 static inline int __add_inode_ref(struct btrfs_trans_handle *trans,
 800                                   struct btrfs_root *root,
 801                                   struct btrfs_path *path,
 802                                   struct btrfs_root *log_root,
 803                                   struct inode *dir, struct inode *inode,
 804                                   struct extent_buffer *eb,
 805                                   u64 inode_objectid, u64 parent_objectid,
 806                                   u64 ref_index, char *name, int namelen,
 807                                   int *search_done)
 808 {
 809         int ret;
 810         char *victim_name;
 811         int victim_name_len;
 812         struct extent_buffer *leaf;
 813         struct btrfs_dir_item *di;
 814         struct btrfs_key search_key;
 815         struct btrfs_inode_extref *extref;
 816
 817 again:
 818         /* Search old style refs */
 819         search_key.objectid = inode_objectid;
 820         search_key.type = BTRFS_INODE_REF_KEY;
 821         search_key.offset = parent_objectid;
 822         ret = btrfs_search_slot(NULL, root, &search_key, path, 0, 0);
 823         if (ret == 0) {
 824                 struct btrfs_inode_ref *victim_ref;
 825                 unsigned long ptr;
 826                 unsigned long ptr_end;
 827
 828                 leaf = path->nodes[0];
 829
 830                 /* are we trying to overwrite a back ref for the root directory
 831                  * if so, just jump out, we're done
 832                  */
 833                 if (search_key.objectid == search_key.offset)
 834                         return 1;
 835
 836                 /* check all the names in this back reference to see
 837                  * if they are in the log.  if so, we allow them to stay
 838                  * otherwise they must be unlinked as a conflict
 839                  */
 840                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
 841                 ptr_end = ptr + btrfs_item_size_nr(leaf, path->slots[0]);
 842                 while (ptr < ptr_end) {
 843                         victim_ref = (struct btrfs_inode_ref *)ptr;
 844                         victim_name_len = btrfs_inode_ref_name_len(leaf,
 845                                                                    victim_ref);
 846                         victim_name = kmalloc(victim_name_len, GFP_NOFS);
 847                         BUG_ON(!victim_name);
 848
 849                         read_extent_buffer(leaf, victim_name,
 850                                            (unsigned long)(victim_ref + 1),
 851                                            victim_name_len);
 852
 853                         if (!backref_in_log(log_root, &search_key,
 854                                             parent_objectid,
 855                                             victim_name,
 856                                             victim_name_len)) {
 857                                 btrfs_inc_nlink(inode);
 858                                 btrfs_release_path(path);
 859
 860                                 ret = btrfs_unlink_inode(trans, root, dir,
 861                                                          inode, victim_name,
 862                                                          victim_name_len);
 863                                 BUG_ON(ret);
 864                                 btrfs_run_delayed_items(trans, root);
 865                                 kfree(victim_name);
 866                                 *search_done = 1;
 867                                 goto again;
 868                         }
 869                         kfree(victim_name);
 870
 871                         ptr = (unsigned long)(victim_ref + 1) + victim_name_len;
 872                 }
 873                 BUG_ON(ret);
 874
 875                 /*
 876                  * NOTE: we have searched root tree and checked the
 877                  * coresponding ref, it does not need to check again.
 878                  */
 879                 *search_done = 1;
 880         }
 881         btrfs_release_path(path);
 882
 883         /* Same search but for extended refs */
 884         extref = btrfs_lookup_inode_extref(NULL, root, path, name, namelen,
 885                                            inode_objectid, parent_objectid, 0,
 886                                            0);
 887         if (!IS_ERR_OR_NULL(extref)) {
 888                 u32 item_size;
 889                 u32 cur_offset = 0;
 890                 unsigned long base;
 891                 struct inode *victim_parent;
 892
 893                 leaf = path->nodes[0];
 894
 895                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 896                 base = btrfs_item_ptr_offset(leaf, path->slots[0]);
 897
 898                 while (cur_offset < item_size) {
 899                         extref = (struct btrfs_inode_extref *)base + cur_offset;
 900
 901                         victim_name_len = btrfs_inode_extref_name_len(leaf, extref);
 902
 903                         if (btrfs_inode_extref_parent(leaf, extref) != parent_objectid)
 904                                 goto next;
 905
 906                         victim_name = kmalloc(victim_name_len, GFP_NOFS);
 907                         read_extent_buffer(leaf, victim_name, (unsigned long)&extref->name,
 908                                            victim_name_len);
 909
 910                         search_key.objectid = inode_objectid;
 911                         search_key.type = BTRFS_INODE_EXTREF_KEY;
 912                         search_key.offset = btrfs_extref_hash(parent_objectid,
 913                                                               victim_name,
 914                                                               victim_name_len);
 915                         ret = 0;
 916                         if (!backref_in_log(log_root, &search_key,
 917                                             parent_objectid, victim_name,
 918                                             victim_name_len)) {
 919                                 ret = -ENOENT;
 920                                 victim_parent = read_one_inode(root,
 921                                                                parent_objectid);
 922                                 if (victim_parent) {
 923                                         btrfs_inc_nlink(inode);
 924                                         btrfs_release_path(path);
 925
 926                                         ret = btrfs_unlink_inode(trans, root,
 927                                                                  victim_parent,
 928                                                                  inode,
 929                                                                  victim_name,
 930                                                                  victim_name_len);
 931                                         btrfs_run_delayed_items(trans, root);
 932                                 }
 933                                 BUG_ON(ret);
 934                                 iput(victim_parent);
 935                                 kfree(victim_name);
 936                                 *search_done = 1;
 937                                 goto again;
 938                         }
 939                         kfree(victim_name);
 940                         BUG_ON(ret);
 941 next:
 942                         cur_offset += victim_name_len + sizeof(*extref);
 943                 }
 944                 *search_done = 1;
 945         }
 946         btrfs_release_path(path);
 947
 948         /* look for a conflicting sequence number */
 949         di = btrfs_lookup_dir_index_item(trans, root, path, btrfs_ino(dir),
 950                                          ref_index, name, namelen, 0);
 951         if (di && !IS_ERR(di)) {
 952                 ret = drop_one_dir_item(trans, root, path, dir, di);
 953                 BUG_ON(ret);
 954         }
 955         btrfs_release_path(path);
 956
 957         /* look for a conflicing name */
 958         di = btrfs_lookup_dir_item(trans, root, path, btrfs_ino(dir),
 959                                    name, namelen, 0);
 960         if (di && !IS_ERR(di)) {
 961                 ret = drop_one_dir_item(trans, root, path, dir, di);
 962                 BUG_ON(ret);
 963         }
 964         btrfs_release_path(path);
 965
 966         return 0;
 967 }
 968
 969 static int extref_get_fields(struct extent_buffer *eb, unsigned long ref_ptr,
 970                              u32 *namelen, char **name, u64 *index,
 971                              u64 *parent_objectid)
 972 {
 973         struct btrfs_inode_extref *extref;
 974
 975         extref = (struct btrfs_inode_extref *)ref_ptr;
 976
 977         *namelen = btrfs_inode_extref_name_len(eb, extref);
 978         *name = kmalloc(*namelen, GFP_NOFS);
 979         if (*name == NULL)
 980                 return -ENOMEM;
 981
 982         read_extent_buffer(eb, *name, (unsigned long)&extref->name,
 983                            *namelen);
 984
 985         *index = btrfs_inode_extref_index(eb, extref);
 986         if (parent_objectid)
 987                 *parent_objectid = btrfs_inode_extref_parent(eb, extref);
 988
 989         return 0;
 990 }
 991
 992 static int ref_get_fields(struct extent_buffer *eb, unsigned long ref_ptr,
 993                           u32 *namelen, char **name, u64 *index)
 994 {
 995         struct btrfs_inode_ref *ref;
 996
 997         ref = (struct btrfs_inode_ref *)ref_ptr;
 998
 999         *namelen = btrfs_inode_ref_name_len(eb, ref);
1000         *name = kmalloc(*namelen, GFP_NOFS);
1001         if (*name == NULL)
1002                 return -ENOMEM;
1003
1004         read_extent_buffer(eb, *name, (unsigned long)(ref + 1), *namelen);
1005
1006         *index = btrfs_inode_ref_index(eb, ref);
1007
1008         return 0;
1009 }
1010
1011 /*
1012  * replay one inode back reference item found in the log tree.
1013  * eb, slot and key refer to the buffer and key found in the log tree.
1014  * root is the destination we are replaying into, and path is for temp
1015  * use by this function.  (it should be released on return).
1016  */
1017 static noinline int add_inode_ref(struct btrfs_trans_handle *trans,
1018                                   struct btrfs_root *root,
1019                                   struct btrfs_root *log,
1020                                   struct btrfs_path *path,
1021                                   struct extent_buffer *eb, int slot,
1022                                   struct btrfs_key *key)
1023 {
1024         struct inode *dir;
1025         struct inode *inode;
1026         unsigned long ref_ptr;
1027         unsigned long ref_end;
1028         char *name;
1029         int namelen;
1030         int ret;
1031         int search_done = 0;
1032         int log_ref_ver = 0;
1033         u64 parent_objectid;
1034         u64 inode_objectid;
1035         u64 ref_index = 0;
1036         int ref_struct_size;
1037
1038         ref_ptr = btrfs_item_ptr_offset(eb, slot);
1039         ref_end = ref_ptr + btrfs_item_size_nr(eb, slot);
1040
1041         if (key->type == BTRFS_INODE_EXTREF_KEY) {
1042                 struct btrfs_inode_extref *r;
1043
1044                 ref_struct_size = sizeof(struct btrfs_inode_extref);
1045                 log_ref_ver = 1;
1046                 r = (struct btrfs_inode_extref *)ref_ptr;
1047                 parent_objectid = btrfs_inode_extref_parent(eb, r);
1048         } else {
1049                 ref_struct_size = sizeof(struct btrfs_inode_ref);
1050                 parent_objectid = key->offset;
1051         }
1052         inode_objectid = key->objectid;
1053
1054         /*
1055          * it is possible that we didn't log all the parent directories
1056          * for a given inode.  If we don't find the dir, just don't
1057          * copy the back ref in.  The link count fixup code will take
1058          * care of the rest
1059          */
1060         dir = read_one_inode(root, parent_objectid);
1061         if (!dir)
1062                 return -ENOENT;
1063
1064         inode = read_one_inode(root, inode_objectid);
1065         if (!inode) {
1066                 iput(dir);
1067                 return -EIO;
1068         }
1069
1070         while (ref_ptr < ref_end) {
1071                 if (log_ref_ver) {
1072                         ret = extref_get_fields(eb, ref_ptr, &namelen, &name,
1073                                                 &ref_index, &parent_objectid);
1074                         /*
1075                          * parent object can change from one array
1076                          * item to another.
1077                          */
1078                         if (!dir)
1079                                 dir = read_one_inode(root, parent_objectid);
1080                         if (!dir)
1081                                 return -ENOENT;
1082                 } else {
1083                         ret = ref_get_fields(eb, ref_ptr, &namelen, &name,
1084                                              &ref_index);
1085                 }
1086                 if (ret)
1087                         return ret;
1088
1089                 /* if we already have a perfect match, we're done */
1090                 if (!inode_in_dir(root, path, btrfs_ino(dir), btrfs_ino(inode),
1091                                   ref_index, name, namelen)) {
1092                         /*
1093                          * look for a conflicting back reference in the
1094                          * metadata. if we find one we have to unlink that name
1095                          * of the file before we add our new link.  Later on, we
1096                          * overwrite any existing back reference, and we don't
1097                          * want to create dangling pointers in the directory.
1098                          */
1099
1100                         if (!search_done) {
1101                                 ret = __add_inode_ref(trans, root, path, log,
1102                                                       dir, inode, eb,
1103                                                       inode_objectid,
1104                                                       parent_objectid,
1105                                                       ref_index, name, namelen,
1106                                                       &search_done);
1107                                 if (ret == 1)
1108                                         goto out;
1109                                 BUG_ON(ret);
1110                         }
1111
1112                         /* insert our name */
1113                         ret = btrfs_add_link(trans, dir, inode, name, namelen,
1114                                              0, ref_index);
1115                         BUG_ON(ret);
1116
1117                         btrfs_update_inode(trans, root, inode);
1118                 }
1119
1120                 ref_ptr = (unsigned long)(ref_ptr + ref_struct_size) + namelen;
1121                 kfree(name);
1122                 if (log_ref_ver) {
1123                         iput(dir);
1124                         dir = NULL;
1125                 }
1126         }
1127
1128         /* finally write the back reference in the inode */
1129         ret = overwrite_item(trans, root, path, eb, slot, key);
1130         BUG_ON(ret);
1131
1132 out:
1133         btrfs_release_path(path);
1134         iput(dir);
1135         iput(inode);
1136         return 0;
1137 }
1138
1139 static int insert_orphan_item(struct btrfs_trans_handle *trans,
1140                               struct btrfs_root *root, u64 offset)
1141 {
1142         int ret;
1143         ret = btrfs_find_orphan_item(root, offset);
1144         if (ret > 0)
1145                 ret = btrfs_insert_orphan_item(trans, root, offset);
1146         return ret;
1147 }
1148
1149 static int count_inode_extrefs(struct btrfs_root *root,
1150                                struct inode *inode, struct btrfs_path *path)
1151 {
1152         int ret = 0;
1153         int name_len;
1154         unsigned int nlink = 0;
1155         u32 item_size;
1156         u32 cur_offset = 0;
1157         u64 inode_objectid = btrfs_ino(inode);
1158         u64 offset = 0;
1159         unsigned long ptr;
1160         struct btrfs_inode_extref *extref;
1161         struct extent_buffer *leaf;
1162
1163         while (1) {
1164                 ret = btrfs_find_one_extref(root, inode_objectid, offset, path,
1165                                             &extref, &offset);
1166                 if (ret)
1167                         break;
1168
1169                 leaf = path->nodes[0];
1170                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1171                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
1172
1173                 while (cur_offset < item_size) {
1174                         extref = (struct btrfs_inode_extref *) (ptr + cur_offset);
1175                         name_len = btrfs_inode_extref_name_len(leaf, extref);
1176
1177                         nlink++;
1178
1179                         cur_offset += name_len + sizeof(*extref);
1180                 }
1181
1182                 offset++;
1183                 btrfs_release_path(path);
1184         }
1185         btrfs_release_path(path);
1186
1187         if (ret < 0)
1188                 return ret;
1189         return nlink;
1190 }
1191
1192 static int count_inode_refs(struct btrfs_root *root,
1193                                struct inode *inode, struct btrfs_path *path)
1194 {
1195         int ret;
1196         struct btrfs_key key;
1197         unsigned int nlink = 0;
1198         unsigned long ptr;
1199         unsigned long ptr_end;
1200         int name_len;
1201         u64 ino = btrfs_ino(inode);
1202
1203         key.objectid = ino;
1204         key.type = BTRFS_INODE_REF_KEY;
1205         key.offset = (u64)-1;
1206
1207         while (1) {
1208                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1209                 if (ret < 0)
1210                         break;
1211                 if (ret > 0) {
1212                         if (path->slots[0] == 0)
1213                                 break;
1214                         path->slots[0]--;
1215                 }
1216                 btrfs_item_key_to_cpu(path->nodes[0], &key,
1217                                       path->slots[0]);
1218                 if (key.objectid != ino ||
1219                     key.type != BTRFS_INODE_REF_KEY)
1220                         break;
1221                 ptr = btrfs_item_ptr_offset(path->nodes[0], path->slots[0]);
1222                 ptr_end = ptr + btrfs_item_size_nr(path->nodes[0],
1223                                                    path->slots[0]);
1224                 while (ptr < ptr_end) {
1225                         struct btrfs_inode_ref *ref;
1226
1227                         ref = (struct btrfs_inode_ref *)ptr;
1228                         name_len = btrfs_inode_ref_name_len(path->nodes[0],
1229                                                             ref);
1230                         ptr = (unsigned long)(ref + 1) + name_len;
1231                         nlink++;
1232                 }
1233
1234                 if (key.offset == 0)
1235                         break;
1236                 key.offset--;
1237                 btrfs_release_path(path);
1238         }
1239         btrfs_release_path(path);
1240
1241         return nlink;
1242 }
1243
1244 /*
1245  * There are a few corners where the link count of the file can't
1246  * be properly maintained during replay.  So, instead of adding
1247  * lots of complexity to the log code, we just scan the backrefs
1248  * for any file that has been through replay.
1249  *
1250  * The scan will update the link count on the inode to reflect the
1251  * number of back refs found.  If it goes down to zero, the iput
1252  * will free the inode.
1253  */
1254 static noinline int fixup_inode_link_count(struct btrfs_trans_handle *trans,
1255                                            struct btrfs_root *root,
1256                                            struct inode *inode)
1257 {
1258         struct btrfs_path *path;
1259         int ret;
1260         u64 nlink = 0;
1261         u64 ino = btrfs_ino(inode);
1262
1263         path = btrfs_alloc_path();
1264         if (!path)
1265                 return -ENOMEM;
1266
1267         ret = count_inode_refs(root, inode, path);
1268         if (ret < 0)
1269                 goto out;
1270
1271         nlink = ret;
1272
1273         ret = count_inode_extrefs(root, inode, path);
1274         if (ret == -ENOENT)
1275                 ret = 0;
1276
1277         if (ret < 0)
1278                 goto out;
1279
1280         nlink += ret;
1281
1282         ret = 0;
1283
1284         if (nlink != inode->i_nlink) {
1285                 set_nlink(inode, nlink);
1286                 btrfs_update_inode(trans, root, inode);
1287         }
1288         BTRFS_I(inode)->index_cnt = (u64)-1;
1289
1290         if (inode->i_nlink == 0) {
1291                 if (S_ISDIR(inode->i_mode)) {
1292                         ret = replay_dir_deletes(trans, root, NULL, path,
1293                                                  ino, 1);
1294                         BUG_ON(ret);
1295                 }
1296                 ret = insert_orphan_item(trans, root, ino);
1297                 BUG_ON(ret);
1298         }
1299
1300 out:
1301         btrfs_free_path(path);
1302         return ret;
1303 }
1304
1305 static noinline int fixup_inode_link_counts(struct btrfs_trans_handle *trans,
1306                                             struct btrfs_root *root,
1307                                             struct btrfs_path *path)
1308 {
1309         int ret;
1310         struct btrfs_key key;
1311         struct inode *inode;
1312
1313         key.objectid = BTRFS_TREE_LOG_FIXUP_OBJECTID;
1314         key.type = BTRFS_ORPHAN_ITEM_KEY;
1315         key.offset = (u64)-1;
1316         while (1) {
1317                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1318                 if (ret < 0)
1319                         break;
1320
1321                 if (ret == 1) {
1322                         if (path->slots[0] == 0)
1323                                 break;
1324                         path->slots[0]--;
1325                 }
1326
1327                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1328                 if (key.objectid != BTRFS_TREE_LOG_FIXUP_OBJECTID ||
1329                     key.type != BTRFS_ORPHAN_ITEM_KEY)
1330                         break;
1331
1332                 ret = btrfs_del_item(trans, root, path);
1333                 if (ret)
1334                         goto out;
1335
1336                 btrfs_release_path(path);
1337                 inode = read_one_inode(root, key.offset);
1338                 if (!inode)
1339                         return -EIO;
1340
1341                 ret = fixup_inode_link_count(trans, root, inode);
1342                 BUG_ON(ret);
1343
1344                 iput(inode);
1345
1346                 /*
1347                  * fixup on a directory may create new entries,
1348                  * make sure we always look for the highset possible
1349                  * offset
1350                  */
1351                 key.offset = (u64)-1;
1352         }
1353         ret = 0;
1354 out:
1355         btrfs_release_path(path);
1356         return ret;
1357 }
1358
1359
1360 /*
1361  * record a given inode in the fixup dir so we can check its link
1362  * count when replay is done.  The link count is incremented here
1363  * so the inode won't go away until we check it
1364  */
1365 static noinline int link_to_fixup_dir(struct btrfs_trans_handle *trans,
1366                                       struct btrfs_root *root,
1367                                       struct btrfs_path *path,
1368                                       u64 objectid)
1369 {
1370         struct btrfs_key key;
1371         int ret = 0;
1372         struct inode *inode;
1373
1374         inode = read_one_inode(root, objectid);
1375         if (!inode)
1376                 return -EIO;
1377
1378         key.objectid = BTRFS_TREE_LOG_FIXUP_OBJECTID;
1379         btrfs_set_key_type(&key, BTRFS_ORPHAN_ITEM_KEY);
1380         key.offset = objectid;
1381
1382         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1383
1384         btrfs_release_path(path);
1385         if (ret == 0) {
1386                 btrfs_inc_nlink(inode);
1387                 ret = btrfs_update_inode(trans, root, inode);
1388         } else if (ret == -EEXIST) {
1389                 ret = 0;
1390         } else {
1391                 BUG();
1392         }
1393         iput(inode);
1394
1395         return ret;
1396 }
1397
1398 /*
1399  * when replaying the log for a directory, we only insert names
1400  * for inodes that actually exist.  This means an fsync on a directory
1401  * does not implicitly fsync all the new files in it
1402  */
1403 static noinline int insert_one_name(struct btrfs_trans_handle *trans,
1404                                     struct btrfs_root *root,
1405                                     struct btrfs_path *path,
1406                                     u64 dirid, u64 index,
1407                                     char *name, int name_len, u8 type,
1408                                     struct btrfs_key *location)
1409 {
1410         struct inode *inode;
1411         struct inode *dir;
1412         int ret;
1413
1414         inode = read_one_inode(root, location->objectid);
1415         if (!inode)
1416                 return -ENOENT;
1417
1418         dir = read_one_inode(root, dirid);
1419         if (!dir) {
1420                 iput(inode);
1421                 return -EIO;
1422         }
1423         ret = btrfs_add_link(trans, dir, inode, name, name_len, 1, index);
1424
1425         /* FIXME, put inode into FIXUP list */
1426
1427         iput(inode);
1428         iput(dir);
1429         return ret;
1430 }
1431
1432 /*
1433  * take a single entry in a log directory item and replay it into
1434  * the subvolume.
1435  *
1436  * if a conflicting item exists in the subdirectory already,
1437  * the inode it points to is unlinked and put into the link count
1438  * fix up tree.
1439  *
1440  * If a name from the log points to a file or directory that does
1441  * not exist in the FS, it is skipped.  fsyncs on directories
1442  * do not force down inodes inside that directory, just changes to the
1443  * names or unlinks in a directory.
1444  */
1445 static noinline int replay_one_name(struct btrfs_trans_handle *trans,
1446                                     struct btrfs_root *root,
1447                                     struct btrfs_path *path,
1448                                     struct extent_buffer *eb,
1449                                     struct btrfs_dir_item *di,
1450                                     struct btrfs_key *key)
1451 {
1452         char *name;
1453         int name_len;
1454         struct btrfs_dir_item *dst_di;
1455         struct btrfs_key found_key;
1456         struct btrfs_key log_key;
1457         struct inode *dir;
1458         u8 log_type;
1459         int exists;
1460         int ret;
1461
1462         dir = read_one_inode(root, key->objectid);
1463         if (!dir)
1464                 return -EIO;
1465
1466         name_len = btrfs_dir_name_len(eb, di);
1467         name = kmalloc(name_len, GFP_NOFS);
1468         if (!name)
1469                 return -ENOMEM;
1470
1471         log_type = btrfs_dir_type(eb, di);
1472         read_extent_buffer(eb, name, (unsigned long)(di + 1),
1473                    name_len);
1474
1475         btrfs_dir_item_key_to_cpu(eb, di, &log_key);
1476         exists = btrfs_lookup_inode(trans, root, path, &log_key, 0);
1477         if (exists == 0)
1478                 exists = 1;
1479         else
1480                 exists = 0;
1481         btrfs_release_path(path);
1482
1483         if (key->type == BTRFS_DIR_ITEM_KEY) {
1484                 dst_di = btrfs_lookup_dir_item(trans, root, path, key->objectid,
1485                                        name, name_len, 1);
1486         } else if (key->type == BTRFS_DIR_INDEX_KEY) {
1487                 dst_di = btrfs_lookup_dir_index_item(trans, root, path,
1488                                                      key->objectid,
1489                                                      key->offset, name,
1490                                                      name_len, 1);
1491         } else {
1492                 BUG();
1493         }
1494         if (IS_ERR_OR_NULL(dst_di)) {
1495                 /* we need a sequence number to insert, so we only
1496                  * do inserts for the BTRFS_DIR_INDEX_KEY types
1497                  */
1498                 if (key->type != BTRFS_DIR_INDEX_KEY)
1499                         goto out;
1500                 goto insert;
1501         }
1502
1503         btrfs_dir_item_key_to_cpu(path->nodes[0], dst_di, &found_key);
1504         /* the existing item matches the logged item */
1505         if (found_key.objectid == log_key.objectid &&
1506             found_key.type == log_key.type &&
1507             found_key.offset == log_key.offset &&
1508             btrfs_dir_type(path->nodes[0], dst_di) == log_type) {
1509                 goto out;
1510         }
1511
1512         /*
1513          * don't drop the conflicting directory entry if the inode
1514          * for the new entry doesn't exist
1515          */
1516         if (!exists)
1517                 goto out;
1518
1519         ret = drop_one_dir_item(trans, root, path, dir, dst_di);
1520         BUG_ON(ret);
1521
1522         if (key->type == BTRFS_DIR_INDEX_KEY)
1523                 goto insert;
1524 out:
1525         btrfs_release_path(path);
1526         kfree(name);
1527         iput(dir);
1528         return 0;
1529
1530 insert:
1531         btrfs_release_path(path);
1532         ret = insert_one_name(trans, root, path, key->objectid, key->offset,
1533                               name, name_len, log_type, &log_key);
1534
1535         BUG_ON(ret && ret != -ENOENT);
1536         goto out;
1537 }
1538
1539 /*
1540  * find all the names in a directory item and reconcile them into
1541  * the subvolume.  Only BTRFS_DIR_ITEM_KEY types will have more than
1542  * one name in a directory item, but the same code gets used for
1543  * both directory index types
1544  */
1545 static noinline int replay_one_dir_item(struct btrfs_trans_handle *trans,
1546                                         struct btrfs_root *root,
1547                                         struct btrfs_path *path,
1548                                         struct extent_buffer *eb, int slot,
1549                                         struct btrfs_key *key)
1550 {
1551         int ret;
1552         u32 item_size = btrfs_item_size_nr(eb, slot);
1553         struct btrfs_dir_item *di;
1554         int name_len;
1555         unsigned long ptr;
1556         unsigned long ptr_end;
1557
1558         ptr = btrfs_item_ptr_offset(eb, slot);
1559         ptr_end = ptr + item_size;
1560         while (ptr < ptr_end) {
1561                 di = (struct btrfs_dir_item *)ptr;
1562                 if (verify_dir_item(root, eb, di))
1563                         return -EIO;
1564                 name_len = btrfs_dir_name_len(eb, di);
1565                 ret = replay_one_name(trans, root, path, eb, di, key);
1566                 BUG_ON(ret);
1567                 ptr = (unsigned long)(di + 1);
1568                 ptr += name_len;
1569         }
1570         return 0;
1571 }
1572
1573 /*
1574  * directory replay has two parts.  There are the standard directory
1575  * items in the log copied from the subvolume, and range items
1576  * created in the log while the subvolume was logged.
1577  *
1578  * The range items tell us which parts of the key space the log
1579  * is authoritative for.  During replay, if a key in the subvolume
1580  * directory is in a logged range item, but not actually in the log
1581  * that means it was deleted from the directory before the fsync
1582  * and should be removed.
1583  */
1584 static noinline int find_dir_range(struct btrfs_root *root,
1585                                    struct btrfs_path *path,
1586                                    u64 dirid, int key_type,
1587                                    u64 *start_ret, u64 *end_ret)
1588 {
1589         struct btrfs_key key;
1590         u64 found_end;
1591         struct btrfs_dir_log_item *item;
1592         int ret;
1593         int nritems;
1594
1595         if (*start_ret == (u64)-1)
1596                 return 1;
1597
1598         key.objectid = dirid;
1599         key.type = key_type;
1600         key.offset = *start_ret;
1601
1602         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
1603         if (ret < 0)
1604                 goto out;
1605         if (ret > 0) {
1606                 if (path->slots[0] == 0)
1607                         goto out;
1608                 path->slots[0]--;
1609         }
1610         if (ret != 0)
1611                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1612
1613         if (key.type != key_type || key.objectid != dirid) {
1614                 ret = 1;
1615                 goto next;
1616         }
1617         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
1618                               struct btrfs_dir_log_item);
1619         found_end = btrfs_dir_log_end(path->nodes[0], item);
1620
1621         if (*start_ret >= key.offset && *start_ret <= found_end) {
1622                 ret = 0;
1623                 *start_ret = key.offset;
1624                 *end_ret = found_end;
1625                 goto out;
1626         }
1627         ret = 1;
1628 next:
1629         /* check the next slot in the tree to see if it is a valid item */
1630         nritems = btrfs_header_nritems(path->nodes[0]);
1631         if (path->slots[0] >= nritems) {
1632                 ret = btrfs_next_leaf(root, path);
1633                 if (ret)
1634                         goto out;
1635         } else {
1636                 path->slots[0]++;
1637         }
1638
1639         btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
1640
1641         if (key.type != key_type || key.objectid != dirid) {
1642                 ret = 1;
1643                 goto out;
1644         }
1645         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
1646                               struct btrfs_dir_log_item);
1647         found_end = btrfs_dir_log_end(path->nodes[0], item);
1648         *start_ret = key.offset;
1649         *end_ret = found_end;
1650         ret = 0;
1651 out:
1652         btrfs_release_path(path);
1653         return ret;
1654 }
1655
1656 /*
1657  * this looks for a given directory item in the log.  If the directory
1658  * item is not in the log, the item is removed and the inode it points
1659  * to is unlinked
1660  */
1661 static noinline int check_item_in_log(struct btrfs_trans_handle *trans,
1662                                       struct btrfs_root *root,
1663                                       struct btrfs_root *log,
1664                                       struct btrfs_path *path,
1665                                       struct btrfs_path *log_path,
1666                                       struct inode *dir,
1667                                       struct btrfs_key *dir_key)
1668 {
1669         int ret;
1670         struct extent_buffer *eb;
1671         int slot;
1672         u32 item_size;
1673         struct btrfs_dir_item *di;
1674         struct btrfs_dir_item *log_di;
1675         int name_len;
1676         unsigned long ptr;
1677         unsigned long ptr_end;
1678         char *name;
1679         struct inode *inode;
1680         struct btrfs_key location;
1681
1682 again:
1683         eb = path->nodes[0];
1684         slot = path->slots[0];
1685         item_size = btrfs_item_size_nr(eb, slot);
1686         ptr = btrfs_item_ptr_offset(eb, slot);
1687         ptr_end = ptr + item_size;
1688         while (ptr < ptr_end) {
1689                 di = (struct btrfs_dir_item *)ptr;
1690                 if (verify_dir_item(root, eb, di)) {
1691                         ret = -EIO;
1692                         goto out;
1693                 }
1694
1695                 name_len = btrfs_dir_name_len(eb, di);
1696                 name = kmalloc(name_len, GFP_NOFS);
1697                 if (!name) {
1698                         ret = -ENOMEM;
1699                         goto out;
1700                 }
1701                 read_extent_buffer(eb, name, (unsigned long)(di + 1),
1702                                   name_len);
1703                 log_di = NULL;
1704                 if (log && dir_key->type == BTRFS_DIR_ITEM_KEY) {
1705                         log_di = btrfs_lookup_dir_item(trans, log, log_path,
1706                                                        dir_key->objectid,
1707                                                        name, name_len, 0);
1708                 } else if (log && dir_key->type == BTRFS_DIR_INDEX_KEY) {
1709                         log_di = btrfs_lookup_dir_index_item(trans, log,
1710                                                      log_path,
1711                                                      dir_key->objectid,
1712                                                      dir_key->offset,
1713                                                      name, name_len, 0);
1714                 }
1715                 if (IS_ERR_OR_NULL(log_di)) {
1716                         btrfs_dir_item_key_to_cpu(eb, di, &location);
1717                         btrfs_release_path(path);
1718                         btrfs_release_path(log_path);
1719                         inode = read_one_inode(root, location.objectid);
1720                         if (!inode) {
1721                                 kfree(name);
1722                                 return -EIO;
1723                         }
1724
1725                         ret = link_to_fixup_dir(trans, root,
1726                                                 path, location.objectid);
1727                         BUG_ON(ret);
1728                         btrfs_inc_nlink(inode);
1729                         ret = btrfs_unlink_inode(trans, root, dir, inode,
1730                                                  name, name_len);
1731                         BUG_ON(ret);
1732
1733                         btrfs_run_delayed_items(trans, root);
1734
1735                         kfree(name);
1736                         iput(inode);
1737
1738                         /* there might still be more names under this key
1739                          * check and repeat if required
1740                          */
1741                         ret = btrfs_search_slot(NULL, root, dir_key, path,
1742                                                 0, 0);
1743                         if (ret == 0)
1744                                 goto again;
1745                         ret = 0;
1746                         goto out;
1747                 }
1748                 btrfs_release_path(log_path);
1749                 kfree(name);
1750
1751                 ptr = (unsigned long)(di + 1);
1752                 ptr += name_len;
1753         }
1754         ret = 0;
1755 out:
1756         btrfs_release_path(path);
1757         btrfs_release_path(log_path);
1758         return ret;
1759 }
1760
1761 /*
1762  * deletion replay happens before we copy any new directory items
1763  * out of the log or out of backreferences from inodes.  It
1764  * scans the log to find ranges of keys that log is authoritative for,
1765  * and then scans the directory to find items in those ranges that are
1766  * not present in the log.
1767  *
1768  * Anything we don't find in the log is unlinked and removed from the
1769  * directory.
1770  */
1771 static noinline int replay_dir_deletes(struct btrfs_trans_handle *trans,
1772                                        struct btrfs_root *root,
1773                                        struct btrfs_root *log,
1774                                        struct btrfs_path *path,
1775                                        u64 dirid, int del_all)
1776 {
1777         u64 range_start;
1778         u64 range_end;
1779         int key_type = BTRFS_DIR_LOG_ITEM_KEY;
1780         int ret = 0;
1781         struct btrfs_key dir_key;
1782         struct btrfs_key found_key;
1783         struct btrfs_path *log_path;
1784         struct inode *dir;
1785
1786         dir_key.objectid = dirid;
1787         dir_key.type = BTRFS_DIR_ITEM_KEY;
1788         log_path = btrfs_alloc_path();
1789         if (!log_path)
1790                 return -ENOMEM;
1791
1792         dir = read_one_inode(root, dirid);
1793         /* it isn't an error if the inode isn't there, that can happen
1794          * because we replay the deletes before we copy in the inode item
1795          * from the log
1796          */
1797         if (!dir) {
1798                 btrfs_free_path(log_path);
1799                 return 0;
1800         }
1801 again:
1802         range_start = 0;
1803         range_end = 0;
1804         while (1) {
1805                 if (del_all)
1806                         range_end = (u64)-1;
1807                 else {
1808                         ret = find_dir_range(log, path, dirid, key_type,
1809                                              &range_start, &range_end);
1810                         if (ret != 0)
1811                                 break;
1812                 }
1813
1814                 dir_key.offset = range_start;
1815                 while (1) {
1816                         int nritems;
1817                         ret = btrfs_search_slot(NULL, root, &dir_key, path,
1818                                                 0, 0);
1819                         if (ret < 0)
1820                                 goto out;
1821
1822                         nritems = btrfs_header_nritems(path->nodes[0]);
1823                         if (path->slots[0] >= nritems) {
1824                                 ret = btrfs_next_leaf(root, path);
1825                                 if (ret)
1826                                         break;
1827                         }
1828                         btrfs_item_key_to_cpu(path->nodes[0], &found_key,
1829                                               path->slots[0]);
1830                         if (found_key.objectid != dirid ||
1831                             found_key.type != dir_key.type)
1832                                 goto next_type;
1833
1834                         if (found_key.offset > range_end)
1835                                 break;
1836
1837                         ret = check_item_in_log(trans, root, log, path,
1838                                                 log_path, dir,
1839                                                 &found_key);
1840                         BUG_ON(ret);
1841                         if (found_key.offset == (u64)-1)
1842                                 break;
1843                         dir_key.offset = found_key.offset + 1;
1844                 }
1845                 btrfs_release_path(path);
1846                 if (range_end == (u64)-1)
1847                         break;
1848                 range_start = range_end + 1;
1849         }
1850
1851 next_type:
1852         ret = 0;
1853         if (key_type == BTRFS_DIR_LOG_ITEM_KEY) {
1854                 key_type = BTRFS_DIR_LOG_INDEX_KEY;
1855                 dir_key.type = BTRFS_DIR_INDEX_KEY;
1856                 btrfs_release_path(path);
1857                 goto again;
1858         }
1859 out:
1860         btrfs_release_path(path);
1861         btrfs_free_path(log_path);
1862         iput(dir);
1863         return ret;
1864 }
1865
1866 /*
1867  * the process_func used to replay items from the log tree.  This
1868  * gets called in two different stages.  The first stage just looks
1869  * for inodes and makes sure they are all copied into the subvolume.
1870  *
1871  * The second stage copies all the other item types from the log into
1872  * the subvolume.  The two stage approach is slower, but gets rid of
1873  * lots of complexity around inodes referencing other inodes that exist
1874  * only in the log (references come from either directory items or inode
1875  * back refs).
1876  */
1877 static int replay_one_buffer(struct btrfs_root *log, struct extent_buffer *eb,
1878                              struct walk_control *wc, u64 gen)
1879 {
1880         int nritems;
1881         struct btrfs_path *path;
1882         struct btrfs_root *root = wc->replay_dest;
1883         struct btrfs_key key;
1884         int level;
1885         int i;
1886         int ret;
1887
1888         ret = btrfs_read_buffer(eb, gen);
1889         if (ret)
1890                 return ret;
1891
1892         level = btrfs_header_level(eb);
1893
1894         if (level != 0)
1895                 return 0;
1896
1897         path = btrfs_alloc_path();
1898         if (!path)
1899                 return -ENOMEM;
1900
1901         nritems = btrfs_header_nritems(eb);
1902         for (i = 0; i < nritems; i++) {
1903                 btrfs_item_key_to_cpu(eb, &key, i);
1904
1905                 /* inode keys are done during the first stage */
1906                 if (key.type == BTRFS_INODE_ITEM_KEY &&
1907                     wc->stage == LOG_WALK_REPLAY_INODES) {
1908                         struct btrfs_inode_item *inode_item;
1909                         u32 mode;
1910
1911                         inode_item = btrfs_item_ptr(eb, i,
1912                                             struct btrfs_inode_item);
1913                         mode = btrfs_inode_mode(eb, inode_item);
1914                         if (S_ISDIR(mode)) {
1915                                 ret = replay_dir_deletes(wc->trans,
1916                                          root, log, path, key.objectid, 0);
1917                                 BUG_ON(ret);
1918                         }
1919                         ret = overwrite_item(wc->trans, root, path,
1920                                              eb, i, &key);
1921                         BUG_ON(ret);
1922
1923                         /* for regular files, make sure corresponding
1924                          * orhpan item exist. extents past the new EOF
1925                          * will be truncated later by orphan cleanup.
1926                          */
1927                         if (S_ISREG(mode)) {
1928                                 ret = insert_orphan_item(wc->trans, root,
1929                                                          key.objectid);
1930                                 BUG_ON(ret);
1931                         }
1932
1933                         ret = link_to_fixup_dir(wc->trans, root,
1934                                                 path, key.objectid);
1935                         BUG_ON(ret);
1936                 }
1937                 if (wc->stage < LOG_WALK_REPLAY_ALL)
1938                         continue;
1939
1940                 /* these keys are simply copied */
1941                 if (key.type == BTRFS_XATTR_ITEM_KEY) {
1942                         ret = overwrite_item(wc->trans, root, path,
1943                                              eb, i, &key);
1944                         BUG_ON(ret);
1945                 } else if (key.type == BTRFS_INODE_REF_KEY) {
1946                         ret = add_inode_ref(wc->trans, root, log, path,
1947                                             eb, i, &key);
1948                         BUG_ON(ret && ret != -ENOENT);
1949                 } else if (key.type == BTRFS_INODE_EXTREF_KEY) {
1950                         ret = add_inode_ref(wc->trans, root, log, path,
1951                                             eb, i, &key);
1952                         BUG_ON(ret && ret != -ENOENT);
1953                 } else if (key.type == BTRFS_EXTENT_DATA_KEY) {
1954                         ret = replay_one_extent(wc->trans, root, path,
1955                                                 eb, i, &key);
1956                         BUG_ON(ret);
1957                 } else if (key.type == BTRFS_DIR_ITEM_KEY ||
1958                            key.type == BTRFS_DIR_INDEX_KEY) {
1959                         ret = replay_one_dir_item(wc->trans, root, path,
1960                                                   eb, i, &key);
1961                         BUG_ON(ret);
1962                 }
1963         }
1964         btrfs_free_path(path);
1965         return 0;
1966 }
1967
1968 static noinline int walk_down_log_tree(struct btrfs_trans_handle *trans,
1969                                    struct btrfs_root *root,
1970                                    struct btrfs_path *path, int *level,
1971                                    struct walk_control *wc)
1972 {
1973         u64 root_owner;
1974         u64 bytenr;
1975         u64 ptr_gen;
1976         struct extent_buffer *next;
1977         struct extent_buffer *cur;
1978         struct extent_buffer *parent;
1979         u32 blocksize;
1980         int ret = 0;
1981
1982         WARN_ON(*level < 0);
1983         WARN_ON(*level >= BTRFS_MAX_LEVEL);
1984
1985         while (*level > 0) {
1986                 WARN_ON(*level < 0);
1987                 WARN_ON(*level >= BTRFS_MAX_LEVEL);
1988                 cur = path->nodes[*level];
1989
1990                 if (btrfs_header_level(cur) != *level)
1991                         WARN_ON(1);
1992
1993                 if (path->slots[*level] >=
1994                     btrfs_header_nritems(cur))
1995                         break;
1996
1997                 bytenr = btrfs_node_blockptr(cur, path->slots[*level]);
1998                 ptr_gen = btrfs_node_ptr_generation(cur, path->slots[*level]);
1999                 blocksize = btrfs_level_size(root, *level - 1);
2000
2001                 parent = path->nodes[*level];
2002                 root_owner = btrfs_header_owner(parent);
2003
2004                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
2005                 if (!next)
2006                         return -ENOMEM;
2007
2008                 if (*level == 1) {
2009                         ret = wc->process_func(root, next, wc, ptr_gen);
2010                         if (ret)
2011                                 return ret;
2012
2013                         path->slots[*level]++;
2014                         if (wc->free) {
2015                                 ret = btrfs_read_buffer(next, ptr_gen);
2016                                 if (ret) {
2017                                         free_extent_buffer(next);
2018                                         return ret;
2019                                 }
2020
2021                                 btrfs_tree_lock(next);
2022                                 btrfs_set_lock_blocking(next);
2023                                 clean_tree_block(trans, root, next);
2024                                 btrfs_wait_tree_block_writeback(next);
2025                                 btrfs_tree_unlock(next);
2026
2027                                 WARN_ON(root_owner !=
2028                                         BTRFS_TREE_LOG_OBJECTID);
2029                                 ret = btrfs_free_and_pin_reserved_extent(root,
2030                                                          bytenr, blocksize);
2031                                 BUG_ON(ret); /* -ENOMEM or logic errors */
2032                         }
2033                         free_extent_buffer(next);
2034                         continue;
2035                 }
2036                 ret = btrfs_read_buffer(next, ptr_gen);
2037                 if (ret) {
2038                         free_extent_buffer(next);
2039                         return ret;
2040                 }
2041
2042                 WARN_ON(*level <= 0);
2043                 if (path->nodes[*level-1])
2044                         free_extent_buffer(path->nodes[*level-1]);
2045                 path->nodes[*level-1] = next;
2046                 *level = btrfs_header_level(next);
2047                 path->slots[*level] = 0;
2048                 cond_resched();
2049         }
2050         WARN_ON(*level < 0);
2051         WARN_ON(*level >= BTRFS_MAX_LEVEL);
2052
2053         path->slots[*level] = btrfs_header_nritems(path->nodes[*level]);
2054
2055         cond_resched();
2056         return 0;
2057 }
2058
2059 static noinline int walk_up_log_tree(struct btrfs_trans_handle *trans,
2060                                  struct btrfs_root *root,
2061                                  struct btrfs_path *path, int *level,
2062                                  struct walk_control *wc)
2063 {
2064         u64 root_owner;
2065         int i;
2066         int slot;
2067         int ret;
2068
2069         for (i = *level; i < BTRFS_MAX_LEVEL - 1 && path->nodes[i]; i++) {
2070                 slot = path->slots[i];
2071                 if (slot + 1 < btrfs_header_nritems(path->nodes[i])) {
2072                         path->slots[i]++;
2073                         *level = i;
2074                         WARN_ON(*level == 0);
2075                         return 0;
2076                 } else {
2077                         struct extent_buffer *parent;
2078                         if (path->nodes[*level] == root->node)
2079                                 parent = path->nodes[*level];
2080                         else
2081                                 parent = path->nodes[*level + 1];
2082
2083                         root_owner = btrfs_header_owner(parent);
2084                         ret = wc->process_func(root, path->nodes[*level], wc,
2085                                  btrfs_header_generation(path->nodes[*level]));
2086                         if (ret)
2087                                 return ret;
2088
2089                         if (wc->free) {
2090                                 struct extent_buffer *next;
2091
2092                                 next = path->nodes[*level];
2093
2094                                 btrfs_tree_lock(next);
2095                                 btrfs_set_lock_blocking(next);
2096                                 clean_tree_block(trans, root, next);
2097                                 btrfs_wait_tree_block_writeback(next);
2098                                 btrfs_tree_unlock(next);
2099
2100                                 WARN_ON(root_owner != BTRFS_TREE_LOG_OBJECTID);
2101                                 ret = btrfs_free_and_pin_reserved_extent(root,
2102                                                 path->nodes[*level]->start,
2103                                                 path->nodes[*level]->len);
2104                                 BUG_ON(ret);
2105                         }
2106                         free_extent_buffer(path->nodes[*level]);
2107                         path->nodes[*level] = NULL;
2108                         *level = i + 1;
2109                 }
2110         }
2111         return 1;
2112 }
2113
2114 /*
2115  * drop the reference count on the tree rooted at 'snap'.  This traverses
2116  * the tree freeing any blocks that have a ref count of zero after being
2117  * decremented.
2118  */
2119 static int walk_log_tree(struct btrfs_trans_handle *trans,
2120                          struct btrfs_root *log, struct walk_control *wc)
2121 {
2122         int ret = 0;
2123         int wret;
2124         int level;
2125         struct btrfs_path *path;
2126         int i;
2127         int orig_level;
2128
2129         path = btrfs_alloc_path();
2130         if (!path)
2131                 return -ENOMEM;
2132
2133         level = btrfs_header_level(log->node);
2134         orig_level = level;
2135         path->nodes[level] = log->node;
2136         extent_buffer_get(log->node);
2137         path->slots[level] = 0;
2138
2139         while (1) {
2140                 wret = walk_down_log_tree(trans, log, path, &level, wc);
2141                 if (wret > 0)
2142                         break;
2143                 if (wret < 0) {
2144                         ret = wret;
2145                         goto out;
2146                 }
2147
2148                 wret = walk_up_log_tree(trans, log, path, &level, wc);
2149                 if (wret > 0)
2150                         break;
2151                 if (wret < 0) {
2152                         ret = wret;
2153                         goto out;
2154                 }
2155         }
2156
2157         /* was the root node processed? if not, catch it here */
2158         if (path->nodes[orig_level]) {
2159                 ret = wc->process_func(log, path->nodes[orig_level], wc,
2160                          btrfs_header_generation(path->nodes[orig_level]));
2161                 if (ret)
2162                         goto out;
2163                 if (wc->free) {
2164                         struct extent_buffer *next;
2165
2166                         next = path->nodes[orig_level];
2167
2168                         btrfs_tree_lock(next);
2169                         btrfs_set_lock_blocking(next);
2170                         clean_tree_block(trans, log, next);
2171                         btrfs_wait_tree_block_writeback(next);
2172                         btrfs_tree_unlock(next);
2173
2174                         WARN_ON(log->root_key.objectid !=
2175                                 BTRFS_TREE_LOG_OBJECTID);
2176                         ret = btrfs_free_and_pin_reserved_extent(log, next->start,
2177                                                          next->len);
2178                         BUG_ON(ret); /* -ENOMEM or logic errors */
2179                 }
2180         }
2181
2182 out:
2183         for (i = 0; i <= orig_level; i++) {
2184                 if (path->nodes[i]) {
2185                         free_extent_buffer(path->nodes[i]);
2186                         path->nodes[i] = NULL;
2187                 }
2188         }
2189         btrfs_free_path(path);
2190         return ret;
2191 }
2192
2193 /*
2194  * helper function to update the item for a given subvolumes log root
2195  * in the tree of log roots
2196  */
2197 static int update_log_root(struct btrfs_trans_handle *trans,
2198                            struct btrfs_root *log)
2199 {
2200         int ret;
2201
2202         if (log->log_transid == 1) {
2203                 /* insert root item on the first sync */
2204                 ret = btrfs_insert_root(trans, log->fs_info->log_root_tree,
2205                                 &log->root_key, &log->root_item);
2206         } else {
2207                 ret = btrfs_update_root(trans, log->fs_info->log_root_tree,
2208                                 &log->root_key, &log->root_item);
2209         }
2210         return ret;
2211 }
2212
2213 static int wait_log_commit(struct btrfs_trans_handle *trans,
2214                            struct btrfs_root *root, unsigned long transid)
2215 {
2216         DEFINE_WAIT(wait);
2217         int index = transid % 2;
2218
2219         /*
2220          * we only allow two pending log transactions at a time,
2221          * so we know that if ours is more than 2 older than the
2222          * current transaction, we're done
2223          */
2224         do {
2225                 prepare_to_wait(&root->log_commit_wait[index],
2226                                 &wait, TASK_UNINTERRUPTIBLE);
2227                 mutex_unlock(&root->log_mutex);
2228
2229                 if (root->fs_info->last_trans_log_full_commit !=
2230                     trans->transid && root->log_transid < transid + 2 &&
2231                     atomic_read(&root->log_commit[index]))
2232                         schedule();
2233
2234                 finish_wait(&root->log_commit_wait[index], &wait);
2235                 mutex_lock(&root->log_mutex);
2236         } while (root->fs_info->last_trans_log_full_commit !=
2237                  trans->transid && root->log_transid < transid + 2 &&
2238                  atomic_read(&root->log_commit[index]));
2239         return 0;
2240 }
2241
2242 static void wait_for_writer(struct btrfs_trans_handle *trans,
2243                             struct btrfs_root *root)
2244 {
2245         DEFINE_WAIT(wait);
2246         while (root->fs_info->last_trans_log_full_commit !=
2247                trans->transid && atomic_read(&root->log_writers)) {
2248                 prepare_to_wait(&root->log_writer_wait,
2249                                 &wait, TASK_UNINTERRUPTIBLE);
2250                 mutex_unlock(&root->log_mutex);
2251                 if (root->fs_info->last_trans_log_full_commit !=
2252                     trans->transid && atomic_read(&root->log_writers))
2253                         schedule();
2254                 mutex_lock(&root->log_mutex);
2255                 finish_wait(&root->log_writer_wait, &wait);
2256         }
2257 }
2258
2259 /*
2260  * btrfs_sync_log does sends a given tree log down to the disk and
2261  * updates the super blocks to record it.  When this call is done,
2262  * you know that any inodes previously logged are safely on disk only
2263  * if it returns 0.
2264  *
2265  * Any other return value means you need to call btrfs_commit_transaction.
2266  * Some of the edge cases for fsyncing directories that have had unlinks
2267  * or renames done in the past mean that sometimes the only safe
2268  * fsync is to commit the whole FS.  When btrfs_sync_log returns -EAGAIN,
2269  * that has happened.
2270  */
2271 int btrfs_sync_log(struct btrfs_trans_handle *trans,
2272                    struct btrfs_root *root)
2273 {
2274         int index1;
2275         int index2;
2276         int mark;
2277         int ret;
2278         struct btrfs_root *log = root->log_root;
2279         struct btrfs_root *log_root_tree = root->fs_info->log_root_tree;
2280         unsigned long log_transid = 0;
2281
2282         mutex_lock(&root->log_mutex);
2283         log_transid = root->log_transid;
2284         index1 = root->log_transid % 2;
2285         if (atomic_read(&root->log_commit[index1])) {
2286                 wait_log_commit(trans, root, root->log_transid);
2287                 mutex_unlock(&root->log_mutex);
2288                 return 0;
2289         }
2290         atomic_set(&root->log_commit[index1], 1);
2291
2292         /* wait for previous tree log sync to complete */
2293         if (atomic_read(&root->log_commit[(index1 + 1) % 2]))
2294                 wait_log_commit(trans, root, root->log_transid - 1);
2295         while (1) {
2296                 int batch = atomic_read(&root->log_batch);
2297                 /* when we're on an ssd, just kick the log commit out */
2298                 if (!btrfs_test_opt(root, SSD) && root->log_multiple_pids) {
2299                         mutex_unlock(&root->log_mutex);
2300                         schedule_timeout_uninterruptible(1);
2301                         mutex_lock(&root->log_mutex);
2302                 }
2303                 wait_for_writer(trans, root);
2304                 if (batch == atomic_read(&root->log_batch))
2305                         break;
2306         }
2307
2308         /* bail out if we need to do a full commit */
2309         if (root->fs_info->last_trans_log_full_commit == trans->transid) {
2310                 ret = -EAGAIN;
2311                 btrfs_free_logged_extents(log, log_transid);
2312                 mutex_unlock(&root->log_mutex);
2313                 goto out;
2314         }
2315
2316         if (log_transid % 2 == 0)
2317                 mark = EXTENT_DIRTY;
2318         else
2319                 mark = EXTENT_NEW;
2320
2321         /* we start IO on  all the marked extents here, but we don't actually
2322          * wait for them until later.
2323          */
2324         ret = btrfs_write_marked_extents(log, &log->dirty_log_pages, mark);
2325         if (ret) {
2326                 btrfs_abort_transaction(trans, root, ret);
2327                 btrfs_free_logged_extents(log, log_transid);
2328                 mutex_unlock(&root->log_mutex);
2329                 goto out;
2330         }
2331
2332         btrfs_set_root_node(&log->root_item, log->node);
2333
2334         root->log_transid++;
2335         log->log_transid = root->log_transid;
2336         root->log_start_pid = 0;
2337         smp_mb();
2338         /*
2339          * IO has been started, blocks of the log tree have WRITTEN flag set
2340          * in their headers. new modifications of the log will be written to
2341          * new positions. so it's safe to allow log writers to go in.
2342          */
2343         mutex_unlock(&root->log_mutex);
2344
2345         mutex_lock(&log_root_tree->log_mutex);
2346         atomic_inc(&log_root_tree->log_batch);
2347         atomic_inc(&log_root_tree->log_writers);
2348         mutex_unlock(&log_root_tree->log_mutex);
2349
2350         ret = update_log_root(trans, log);
2351
2352         mutex_lock(&log_root_tree->log_mutex);
2353         if (atomic_dec_and_test(&log_root_tree->log_writers)) {
2354                 smp_mb();
2355                 if (waitqueue_active(&log_root_tree->log_writer_wait))
2356                         wake_up(&log_root_tree->log_writer_wait);
2357         }
2358
2359         if (ret) {
2360                 if (ret != -ENOSPC) {
2361                         btrfs_abort_transaction(trans, root, ret);
2362                         mutex_unlock(&log_root_tree->log_mutex);
2363                         goto out;
2364                 }
2365                 root->fs_info->last_trans_log_full_commit = trans->transid;
2366                 btrfs_wait_marked_extents(log, &log->dirty_log_pages, mark);
2367                 btrfs_free_logged_extents(log, log_transid);
2368                 mutex_unlock(&log_root_tree->log_mutex);
2369                 ret = -EAGAIN;
2370                 goto out;
2371         }
2372
2373         index2 = log_root_tree->log_transid % 2;
2374         if (atomic_read(&log_root_tree->log_commit[index2])) {
2375                 btrfs_wait_marked_extents(log, &log->dirty_log_pages, mark);
2376                 wait_log_commit(trans, log_root_tree,
2377                                 log_root_tree->log_transid);
2378                 btrfs_free_logged_extents(log, log_transid);
2379                 mutex_unlock(&log_root_tree->log_mutex);
2380                 ret = 0;
2381                 goto out;
2382         }
2383         atomic_set(&log_root_tree->log_commit[index2], 1);
2384
2385         if (atomic_read(&log_root_tree->log_commit[(index2 + 1) % 2])) {
2386                 wait_log_commit(trans, log_root_tree,
2387                                 log_root_tree->log_transid - 1);
2388         }
2389
2390         wait_for_writer(trans, log_root_tree);
2391
2392         /*
2393          * now that we've moved on to the tree of log tree roots,
2394          * check the full commit flag again
2395          */
2396         if (root->fs_info->last_trans_log_full_commit == trans->transid) {
2397                 btrfs_wait_marked_extents(log, &log->dirty_log_pages, mark);
2398                 btrfs_free_logged_extents(log, log_transid);
2399                 mutex_unlock(&log_root_tree->log_mutex);
2400                 ret = -EAGAIN;
2401                 goto out_wake_log_root;
2402         }
2403
2404         ret = btrfs_write_and_wait_marked_extents(log_root_tree,
2405                                 &log_root_tree->dirty_log_pages,
2406                                 EXTENT_DIRTY | EXTENT_NEW);
2407         if (ret) {
2408                 btrfs_abort_transaction(trans, root, ret);
2409                 btrfs_free_logged_extents(log, log_transid);
2410                 mutex_unlock(&log_root_tree->log_mutex);
2411                 goto out_wake_log_root;
2412         }
2413         btrfs_wait_marked_extents(log, &log->dirty_log_pages, mark);
2414         btrfs_wait_logged_extents(log, log_transid);
2415
2416         btrfs_set_super_log_root(root->fs_info->super_for_commit,
2417                                 log_root_tree->node->start);
2418         btrfs_set_super_log_root_level(root->fs_info->super_for_commit,
2419                                 btrfs_header_level(log_root_tree->node));
2420
2421         log_root_tree->log_transid++;
2422         smp_mb();
2423
2424         mutex_unlock(&log_root_tree->log_mutex);
2425
2426         /*
2427          * nobody else is going to jump in and write the the ctree
2428          * super here because the log_commit atomic below is protecting
2429          * us.  We must be called with a transaction handle pinning
2430          * the running transaction open, so a full commit can't hop
2431          * in and cause problems either.
2432          */
2433         btrfs_scrub_pause_super(root);
2434         ret = write_ctree_super(trans, root->fs_info->tree_root, 1);
2435         btrfs_scrub_continue_super(root);
2436         if (ret) {
2437                 btrfs_abort_transaction(trans, root, ret);
2438                 goto out_wake_log_root;
2439         }
2440
2441         mutex_lock(&root->log_mutex);
2442         if (root->last_log_commit < log_transid)
2443                 root->last_log_commit = log_transid;
2444         mutex_unlock(&root->log_mutex);
2445
2446 out_wake_log_root:
2447         atomic_set(&log_root_tree->log_commit[index2], 0);
2448         smp_mb();
2449         if (waitqueue_active(&log_root_tree->log_commit_wait[index2]))
2450                 wake_up(&log_root_tree->log_commit_wait[index2]);
2451 out:
2452         atomic_set(&root->log_commit[index1], 0);
2453         smp_mb();
2454         if (waitqueue_active(&root->log_commit_wait[index1]))
2455                 wake_up(&root->log_commit_wait[index1]);
2456         return ret;
2457 }
2458
2459 static void free_log_tree(struct btrfs_trans_handle *trans,
2460                           struct btrfs_root *log)
2461 {
2462         int ret;
2463         u64 start;
2464         u64 end;
2465         struct walk_control wc = {
2466                 .free = 1,
2467                 .process_func = process_one_buffer
2468         };
2469
2470         ret = walk_log_tree(trans, log, &wc);
2471         BUG_ON(ret);
2472
2473         while (1) {
2474                 ret = find_first_extent_bit(&log->dirty_log_pages,
2475                                 0, &start, &end, EXTENT_DIRTY | EXTENT_NEW,
2476                                 NULL);
2477                 if (ret)
2478                         break;
2479
2480                 clear_extent_bits(&log->dirty_log_pages, start, end,
2481                                   EXTENT_DIRTY | EXTENT_NEW, GFP_NOFS);
2482         }
2483
2484         /*
2485          * We may have short-circuited the log tree with the full commit logic
2486          * and left ordered extents on our list, so clear these out to keep us
2487          * from leaking inodes and memory.
2488          */
2489         btrfs_free_logged_extents(log, 0);
2490         btrfs_free_logged_extents(log, 1);
2491
2492         free_extent_buffer(log->node);
2493         kfree(log);
2494 }
2495
2496 /*
2497  * free all the extents used by the tree log.  This should be called
2498  * at commit time of the full transaction
2499  */
2500 int btrfs_free_log(struct btrfs_trans_handle *trans, struct btrfs_root *root)
2501 {
2502         if (root->log_root) {
2503                 free_log_tree(trans, root->log_root);
2504                 root->log_root = NULL;
2505         }
2506         return 0;
2507 }
2508
2509 int btrfs_free_log_root_tree(struct btrfs_trans_handle *trans,
2510                              struct btrfs_fs_info *fs_info)
2511 {
2512         if (fs_info->log_root_tree) {
2513                 free_log_tree(trans, fs_info->log_root_tree);
2514                 fs_info->log_root_tree = NULL;
2515         }
2516         return 0;
2517 }
2518
2519 /*
2520  * If both a file and directory are logged, and unlinks or renames are
2521  * mixed in, we have a few interesting corners:
2522  *
2523  * create file X in dir Y
2524  * link file X to X.link in dir Y
2525  * fsync file X
2526  * unlink file X but leave X.link
2527  * fsync dir Y
2528  *
2529  * After a crash we would expect only X.link to exist.  But file X
2530  * didn't get fsync'd again so the log has back refs for X and X.link.
2531  *
2532  * We solve this by removing directory entries and inode backrefs from the
2533  * log when a file that was logged in the current transaction is
2534  * unlinked.  Any later fsync will include the updated log entries, and
2535  * we'll be able to reconstruct the proper directory items from backrefs.
2536  *
2537  * This optimizations allows us to avoid relogging the entire inode
2538  * or the entire directory.
2539  */
2540 int btrfs_del_dir_entries_in_log(struct btrfs_trans_handle *trans,
2541                                  struct btrfs_root *root,
2542                                  const char *name, int name_len,
2543                                  struct inode *dir, u64 index)
2544 {
2545         struct btrfs_root *log;
2546         struct btrfs_dir_item *di;
2547         struct btrfs_path *path;
2548         int ret;
2549         int err = 0;
2550         int bytes_del = 0;
2551         u64 dir_ino = btrfs_ino(dir);
2552
2553         if (BTRFS_I(dir)->logged_trans < trans->transid)
2554                 return 0;
2555
2556         ret = join_running_log_trans(root);
2557         if (ret)
2558                 return 0;
2559
2560         mutex_lock(&BTRFS_I(dir)->log_mutex);
2561
2562         log = root->log_root;
2563         path = btrfs_alloc_path();
2564         if (!path) {
2565                 err = -ENOMEM;
2566                 goto out_unlock;
2567         }
2568
2569         di = btrfs_lookup_dir_item(trans, log, path, dir_ino,
2570                                    name, name_len, -1);
2571         if (IS_ERR(di)) {
2572                 err = PTR_ERR(di);
2573                 goto fail;
2574         }
2575         if (di) {
2576                 ret = btrfs_delete_one_dir_name(trans, log, path, di);
2577                 bytes_del += name_len;
2578                 BUG_ON(ret);
2579         }
2580         btrfs_release_path(path);
2581         di = btrfs_lookup_dir_index_item(trans, log, path, dir_ino,
2582                                          index, name, name_len, -1);
2583         if (IS_ERR(di)) {
2584                 err = PTR_ERR(di);
2585                 goto fail;
2586         }
2587         if (di) {
2588                 ret = btrfs_delete_one_dir_name(trans, log, path, di);
2589                 bytes_del += name_len;
2590                 BUG_ON(ret);
2591         }
2592
2593         /* update the directory size in the log to reflect the names
2594          * we have removed
2595          */
2596         if (bytes_del) {
2597                 struct btrfs_key key;
2598
2599                 key.objectid = dir_ino;
2600                 key.offset = 0;
2601                 key.type = BTRFS_INODE_ITEM_KEY;
2602                 btrfs_release_path(path);
2603
2604                 ret = btrfs_search_slot(trans, log, &key, path, 0, 1);
2605                 if (ret < 0) {
2606                         err = ret;
2607                         goto fail;
2608                 }
2609                 if (ret == 0) {
2610                         struct btrfs_inode_item *item;
2611                         u64 i_size;
2612
2613                         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
2614                                               struct btrfs_inode_item);
2615                         i_size = btrfs_inode_size(path->nodes[0], item);
2616                         if (i_size > bytes_del)
2617                                 i_size -= bytes_del;
2618                         else
2619                                 i_size = 0;
2620                         btrfs_set_inode_size(path->nodes[0], item, i_size);
2621                         btrfs_mark_buffer_dirty(path->nodes[0]);
2622                 } else
2623                         ret = 0;
2624                 btrfs_release_path(path);
2625         }
2626 fail:
2627         btrfs_free_path(path);
2628 out_unlock:
2629         mutex_unlock(&BTRFS_I(dir)->log_mutex);
2630         if (ret == -ENOSPC) {
2631                 root->fs_info->last_trans_log_full_commit = trans->transid;
2632                 ret = 0;
2633         } else if (ret < 0)
2634                 btrfs_abort_transaction(trans, root, ret);
2635
2636         btrfs_end_log_trans(root);
2637
2638         return err;
2639 }
2640
2641 /* see comments for btrfs_del_dir_entries_in_log */
2642 int btrfs_del_inode_ref_in_log(struct btrfs_trans_handle *trans,
2643                                struct btrfs_root *root,
2644                                const char *name, int name_len,
2645                                struct inode *inode, u64 dirid)
2646 {
2647         struct btrfs_root *log;
2648         u64 index;
2649         int ret;
2650
2651         if (BTRFS_I(inode)->logged_trans < trans->transid)
2652                 return 0;
2653
2654         ret = join_running_log_trans(root);
2655         if (ret)
2656                 return 0;
2657         log = root->log_root;
2658         mutex_lock(&BTRFS_I(inode)->log_mutex);
2659
2660         ret = btrfs_del_inode_ref(trans, log, name, name_len, btrfs_ino(inode),
2661                                   dirid, &index);
2662         mutex_unlock(&BTRFS_I(inode)->log_mutex);
2663         if (ret == -ENOSPC) {
2664                 root->fs_info->last_trans_log_full_commit = trans->transid;
2665                 ret = 0;
2666         } else if (ret < 0 && ret != -ENOENT)
2667                 btrfs_abort_transaction(trans, root, ret);
2668         btrfs_end_log_trans(root);
2669
2670         return ret;
2671 }
2672
2673 /*
2674  * creates a range item in the log for 'dirid'.  first_offset and
2675  * last_offset tell us which parts of the key space the log should
2676  * be considered authoritative for.
2677  */
2678 static noinline int insert_dir_log_key(struct btrfs_trans_handle *trans,
2679                                        struct btrfs_root *log,
2680                                        struct btrfs_path *path,
2681                                        int key_type, u64 dirid,
2682                                        u64 first_offset, u64 last_offset)
2683 {
2684         int ret;
2685         struct btrfs_key key;
2686         struct btrfs_dir_log_item *item;
2687
2688         key.objectid = dirid;
2689         key.offset = first_offset;
2690         if (key_type == BTRFS_DIR_ITEM_KEY)
2691                 key.type = BTRFS_DIR_LOG_ITEM_KEY;
2692         else
2693                 key.type = BTRFS_DIR_LOG_INDEX_KEY;
2694         ret = btrfs_insert_empty_item(trans, log, path, &key, sizeof(*item));
2695         if (ret)
2696                 return ret;
2697
2698         item = btrfs_item_ptr(path->nodes[0], path->slots[0],
2699                               struct btrfs_dir_log_item);
2700         btrfs_set_dir_log_end(path->nodes[0], item, last_offset);
2701         btrfs_mark_buffer_dirty(path->nodes[0]);
2702         btrfs_release_path(path);
2703         return 0;
2704 }
2705
2706 /*
2707  * log all the items included in the current transaction for a given
2708  * directory.  This also creates the range items in the log tree required
2709  * to replay anything deleted before the fsync
2710  */
2711 static noinline int log_dir_items(struct btrfs_trans_handle *trans,
2712                           struct btrfs_root *root, struct inode *inode,
2713                           struct btrfs_path *path,
2714                           struct btrfs_path *dst_path, int key_type,
2715                           u64 min_offset, u64 *last_offset_ret)
2716 {
2717         struct btrfs_key min_key;
2718         struct btrfs_key max_key;
2719         struct btrfs_root *log = root->log_root;
2720         struct extent_buffer *src;
2721         int err = 0;
2722         int ret;
2723         int i;
2724         int nritems;
2725         u64 first_offset = min_offset;
2726         u64 last_offset = (u64)-1;
2727         u64 ino = btrfs_ino(inode);
2728
2729         log = root->log_root;
2730         max_key.objectid = ino;
2731         max_key.offset = (u64)-1;
2732         max_key.type = key_type;
2733
2734         min_key.objectid = ino;
2735         min_key.type = key_type;
2736         min_key.offset = min_offset;
2737
2738         path->keep_locks = 1;
2739
2740         ret = btrfs_search_forward(root, &min_key, &max_key,
2741                                    path, trans->transid);
2742
2743         /*
2744          * we didn't find anything from this transaction, see if there
2745          * is anything at all
2746          */
2747         if (ret != 0 || min_key.objectid != ino || min_key.type != key_type) {
2748                 min_key.objectid = ino;
2749                 min_key.type = key_type;
2750                 min_key.offset = (u64)-1;
2751                 btrfs_release_path(path);
2752                 ret = btrfs_search_slot(NULL, root, &min_key, path, 0, 0);
2753                 if (ret < 0) {
2754                         btrfs_release_path(path);
2755                         return ret;
2756                 }
2757                 ret = btrfs_previous_item(root, path, ino, key_type);
2758
2759                 /* if ret == 0 there are items for this type,
2760                  * create a range to tell us the last key of this type.
2761                  * otherwise, there are no items in this directory after
2762                  * *min_offset, and we create a range to indicate that.
2763                  */
2764                 if (ret == 0) {
2765                         struct btrfs_key tmp;
2766                         btrfs_item_key_to_cpu(path->nodes[0], &tmp,
2767                                               path->slots[0]);
2768                         if (key_type == tmp.type)
2769                                 first_offset = max(min_offset, tmp.offset) + 1;
2770                 }
2771                 goto done;
2772         }
2773
2774         /* go backward to find any previous key */
2775         ret = btrfs_previous_item(root, path, ino, key_type);
2776         if (ret == 0) {
2777                 struct btrfs_key tmp;
2778                 btrfs_item_key_to_cpu(path->nodes[0], &tmp, path->slots[0]);
2779                 if (key_type == tmp.type) {
2780                         first_offset = tmp.offset;
2781                         ret = overwrite_item(trans, log, dst_path,
2782                                              path->nodes[0], path->slots[0],
2783                                              &tmp);
2784                         if (ret) {
2785                                 err = ret;
2786                                 goto done;
2787                         }
2788                 }
2789         }
2790         btrfs_release_path(path);
2791
2792         /* find the first key from this transaction again */
2793         ret = btrfs_search_slot(NULL, root, &min_key, path, 0, 0);
2794         if (ret != 0) {
2795                 WARN_ON(1);
2796                 goto done;
2797         }
2798
2799         /*
2800          * we have a block from this transaction, log every item in it
2801          * from our directory
2802          */
2803         while (1) {
2804                 struct btrfs_key tmp;
2805                 src = path->nodes[0];
2806                 nritems = btrfs_header_nritems(src);
2807                 for (i = path->slots[0]; i < nritems; i++) {
2808                         btrfs_item_key_to_cpu(src, &min_key, i);
2809
2810                         if (min_key.objectid != ino || min_key.type != key_type)
2811                                 goto done;
2812                         ret = overwrite_item(trans, log, dst_path, src, i,
2813                                              &min_key);
2814                         if (ret) {
2815                                 err = ret;
2816                                 goto done;
2817                         }
2818                 }
2819                 path->slots[0] = nritems;
2820
2821                 /*
2822                  * look ahead to the next item and see if it is also
2823                  * from this directory and from this transaction
2824                  */
2825                 ret = btrfs_next_leaf(root, path);
2826                 if (ret == 1) {
2827                         last_offset = (u64)-1;
2828                         goto done;
2829                 }
2830                 btrfs_item_key_to_cpu(path->nodes[0], &tmp, path->slots[0]);
2831                 if (tmp.objectid != ino || tmp.type != key_type) {
2832                         last_offset = (u64)-1;
2833                         goto done;
2834                 }
2835                 if (btrfs_header_generation(path->nodes[0]) != trans->transid) {
2836                         ret = overwrite_item(trans, log, dst_path,
2837                                              path->nodes[0], path->slots[0],
2838                                              &tmp);
2839                         if (ret)
2840                                 err = ret;
2841                         else
2842                                 last_offset = tmp.offset;
2843                         goto done;
2844                 }
2845         }
2846 done:
2847         btrfs_release_path(path);
2848         btrfs_release_path(dst_path);
2849
2850         if (err == 0) {
2851                 *last_offset_ret = last_offset;
2852                 /*
2853                  * insert the log range keys to indicate where the log
2854                  * is valid
2855                  */
2856                 ret = insert_dir_log_key(trans, log, path, key_type,
2857                                          ino, first_offset, last_offset);
2858                 if (ret)
2859                         err = ret;
2860         }
2861         return err;
2862 }
2863
2864 /*
2865  * logging directories is very similar to logging inodes, We find all the items
2866  * from the current transaction and write them to the log.
2867  *
2868  * The recovery code scans the directory in the subvolume, and if it finds a
2869  * key in the range logged that is not present in the log tree, then it means
2870  * that dir entry was unlinked during the transaction.
2871  *
2872  * In order for that scan to work, we must include one key smaller than
2873  * the smallest logged by this transaction and one key larger than the largest
2874  * key logged by this transaction.
2875  */
2876 static noinline int log_directory_changes(struct btrfs_trans_handle *trans,
2877                           struct btrfs_root *root, struct inode *inode,
2878                           struct btrfs_path *path,
2879                           struct btrfs_path *dst_path)
2880 {
2881         u64 min_key;
2882         u64 max_key;
2883         int ret;
2884         int key_type = BTRFS_DIR_ITEM_KEY;
2885
2886 again:
2887         min_key = 0;
2888         max_key = 0;
2889         while (1) {
2890                 ret = log_dir_items(trans, root, inode, path,
2891                                     dst_path, key_type, min_key,
2892                                     &max_key);
2893                 if (ret)
2894                         return ret;
2895                 if (max_key == (u64)-1)
2896                         break;
2897                 min_key = max_key + 1;
2898         }
2899
2900         if (key_type == BTRFS_DIR_ITEM_KEY) {
2901                 key_type = BTRFS_DIR_INDEX_KEY;
2902                 goto again;
2903         }
2904         return 0;
2905 }
2906
2907 /*
2908  * a helper function to drop items from the log before we relog an
2909  * inode.  max_key_type indicates the highest item type to remove.
2910  * This cannot be run for file data extents because it does not
2911  * free the extents they point to.
2912  */
2913 static int drop_objectid_items(struct btrfs_trans_handle *trans,
2914                                   struct btrfs_root *log,
2915                                   struct btrfs_path *path,
2916                                   u64 objectid, int max_key_type)
2917 {
2918         int ret;
2919         struct btrfs_key key;
2920         struct btrfs_key found_key;
2921         int start_slot;
2922
2923         key.objectid = objectid;
2924         key.type = max_key_type;
2925         key.offset = (u64)-1;
2926
2927         while (1) {
2928                 ret = btrfs_search_slot(trans, log, &key, path, -1, 1);
2929                 BUG_ON(ret == 0);
2930                 if (ret < 0)
2931                         break;
2932
2933                 if (path->slots[0] == 0)
2934                         break;
2935
2936                 path->slots[0]--;
2937                 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
2938                                       path->slots[0]);
2939
2940                 if (found_key.objectid != objectid)
2941                         break;
2942
2943                 found_key.offset = 0;
2944                 found_key.type = 0;
2945                 ret = btrfs_bin_search(path->nodes[0], &found_key, 0,
2946                                        &start_slot);
2947
2948                 ret = btrfs_del_items(trans, log, path, start_slot,
2949                                       path->slots[0] - start_slot + 1);
2950                 /*
2951                  * If start slot isn't 0 then we don't need to re-search, we've
2952                  * found the last guy with the objectid in this tree.
2953                  */
2954                 if (ret || start_slot != 0)
2955                         break;
2956                 btrfs_release_path(path);
2957         }
2958         btrfs_release_path(path);
2959         if (ret > 0)
2960                 ret = 0;
2961         return ret;
2962 }
2963
2964 static void fill_inode_item(struct btrfs_trans_handle *trans,
2965                             struct extent_buffer *leaf,
2966                             struct btrfs_inode_item *item,
2967                             struct inode *inode, int log_inode_only)
2968 {
2969         struct btrfs_map_token token;
2970
2971         btrfs_init_map_token(&token);
2972
2973         if (log_inode_only) {
2974                 /* set the generation to zero so the recover code
2975                  * can tell the difference between an logging
2976                  * just to say 'this inode exists' and a logging
2977                  * to say 'update this inode with these values'
2978                  */
2979                 btrfs_set_token_inode_generation(leaf, item, 0, &token);
2980                 btrfs_set_token_inode_size(leaf, item, 0, &token);
2981         } else {
2982                 btrfs_set_token_inode_generation(leaf, item,
2983                                                  BTRFS_I(inode)->generation,
2984                                                  &token);
2985                 btrfs_set_token_inode_size(leaf, item, inode->i_size, &token);
2986         }
2987
2988         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
2989         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
2990         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
2991         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
2992
2993         btrfs_set_token_timespec_sec(leaf, btrfs_inode_atime(item),
2994                                      inode->i_atime.tv_sec, &token);
2995         btrfs_set_token_timespec_nsec(leaf, btrfs_inode_atime(item),
2996                                       inode->i_atime.tv_nsec, &token);
2997
2998         btrfs_set_token_timespec_sec(leaf, btrfs_inode_mtime(item),
2999                                      inode->i_mtime.tv_sec, &token);
3000         btrfs_set_token_timespec_nsec(leaf, btrfs_inode_mtime(item),
3001                                       inode->i_mtime.tv_nsec, &token);
3002
3003         btrfs_set_token_timespec_sec(leaf, btrfs_inode_ctime(item),
3004                                      inode->i_ctime.tv_sec, &token);
3005         btrfs_set_token_timespec_nsec(leaf, btrfs_inode_ctime(item),
3006                                       inode->i_ctime.tv_nsec, &token);
3007
3008         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3009                                      &token);
3010
3011         btrfs_set_token_inode_sequence(leaf, item, inode->i_version, &token);
3012         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3013         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3014         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3015         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3016 }
3017
3018 static int log_inode_item(struct btrfs_trans_handle *trans,
3019                           struct btrfs_root *log, struct btrfs_path *path,
3020                           struct inode *inode)
3021 {
3022         struct btrfs_inode_item *inode_item;
3023         struct btrfs_key key;
3024         int ret;
3025
3026         memcpy(&key, &BTRFS_I(inode)->location, sizeof(key));
3027         ret = btrfs_insert_empty_item(trans, log, path, &key,
3028                                       sizeof(*inode_item));
3029         if (ret && ret != -EEXIST)
3030                 return ret;
3031         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
3032                                     struct btrfs_inode_item);
3033         fill_inode_item(trans, path->nodes[0], inode_item, inode, 0);
3034         btrfs_release_path(path);
3035         return 0;
3036 }
3037
3038 static noinline int copy_items(struct btrfs_trans_handle *trans,
3039                                struct inode *inode,
3040                                struct btrfs_path *dst_path,
3041                                struct extent_buffer *src,
3042                                int start_slot, int nr, int inode_only)
3043 {
3044         unsigned long src_offset;
3045         unsigned long dst_offset;
3046         struct btrfs_root *log = BTRFS_I(inode)->root->log_root;
3047         struct btrfs_file_extent_item *extent;
3048         struct btrfs_inode_item *inode_item;
3049         int ret;
3050         struct btrfs_key *ins_keys;
3051         u32 *ins_sizes;
3052         char *ins_data;
3053         int i;
3054         struct list_head ordered_sums;
3055         int skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
3056
3057         INIT_LIST_HEAD(&ordered_sums);
3058
3059         ins_data = kmalloc(nr * sizeof(struct btrfs_key) +
3060                            nr * sizeof(u32), GFP_NOFS);
3061         if (!ins_data)
3062                 return -ENOMEM;
3063
3064         ins_sizes = (u32 *)ins_data;
3065         ins_keys = (struct btrfs_key *)(ins_data + nr * sizeof(u32));
3066
3067         for (i = 0; i < nr; i++) {
3068                 ins_sizes[i] = btrfs_item_size_nr(src, i + start_slot);
3069                 btrfs_item_key_to_cpu(src, ins_keys + i, i + start_slot);
3070         }
3071         ret = btrfs_insert_empty_items(trans, log, dst_path,
3072                                        ins_keys, ins_sizes, nr);
3073         if (ret) {
3074                 kfree(ins_data);
3075                 return ret;
3076         }
3077
3078         for (i = 0; i < nr; i++, dst_path->slots[0]++) {
3079                 dst_offset = btrfs_item_ptr_offset(dst_path->nodes[0],
3080                                                    dst_path->slots[0]);
3081
3082                 src_offset = btrfs_item_ptr_offset(src, start_slot + i);
3083
3084                 if (ins_keys[i].type == BTRFS_INODE_ITEM_KEY) {
3085                         inode_item = btrfs_item_ptr(dst_path->nodes[0],
3086                                                     dst_path->slots[0],
3087                                                     struct btrfs_inode_item);
3088                         fill_inode_item(trans, dst_path->nodes[0], inode_item,
3089                                         inode, inode_only == LOG_INODE_EXISTS);
3090                 } else {
3091                         copy_extent_buffer(dst_path->nodes[0], src, dst_offset,
3092                                            src_offset, ins_sizes[i]);
3093                 }
3094
3095                 /* take a reference on file data extents so that truncates
3096                  * or deletes of this inode don't have to relog the inode
3097                  * again
3098                  */
3099                 if (btrfs_key_type(ins_keys + i) == BTRFS_EXTENT_DATA_KEY &&
3100                     !skip_csum) {
3101                         int found_type;
3102                         extent = btrfs_item_ptr(src, start_slot + i,
3103                                                 struct btrfs_file_extent_item);
3104
3105                         if (btrfs_file_extent_generation(src, extent) < trans->transid)
3106                                 continue;
3107
3108                         found_type = btrfs_file_extent_type(src, extent);
3109                         if (found_type == BTRFS_FILE_EXTENT_REG) {
3110                                 u64 ds, dl, cs, cl;
3111                                 ds = btrfs_file_extent_disk_bytenr(src,
3112                                                                 extent);
3113                                 /* ds == 0 is a hole */
3114                                 if (ds == 0)
3115                                         continue;
3116
3117                                 dl = btrfs_file_extent_disk_num_bytes(src,
3118                                                                 extent);
3119                                 cs = btrfs_file_extent_offset(src, extent);
3120                                 cl = btrfs_file_extent_num_bytes(src,
3121                                                                 extent);
3122                                 if (btrfs_file_extent_compression(src,
3123                                                                   extent)) {
3124                                         cs = 0;
3125                                         cl = dl;
3126                                 }
3127
3128                                 ret = btrfs_lookup_csums_range(
3129                                                 log->fs_info->csum_root,
3130                                                 ds + cs, ds + cs + cl - 1,
3131                                                 &ordered_sums, 0);
3132                                 BUG_ON(ret);
3133                         }
3134                 }
3135         }
3136
3137         btrfs_mark_buffer_dirty(dst_path->nodes[0]);
3138         btrfs_release_path(dst_path);
3139         kfree(ins_data);
3140
3141         /*
3142          * we have to do this after the loop above to avoid changing the
3143          * log tree while trying to change the log tree.
3144          */
3145         ret = 0;
3146         while (!list_empty(&ordered_sums)) {
3147                 struct btrfs_ordered_sum *sums = list_entry(ordered_sums.next,
3148                                                    struct btrfs_ordered_sum,
3149                                                    list);
3150                 if (!ret)
3151                         ret = btrfs_csum_file_blocks(trans, log, sums);
3152                 list_del(&sums->list);
3153                 kfree(sums);
3154         }
3155         return ret;
3156 }
3157
3158 static int extent_cmp(void *priv, struct list_head *a, struct list_head *b)
3159 {
3160         struct extent_map *em1, *em2;
3161
3162         em1 = list_entry(a, struct extent_map, list);
3163         em2 = list_entry(b, struct extent_map, list);
3164
3165         if (em1->start < em2->start)
3166                 return -1;
3167         else if (em1->start > em2->start)
3168                 return 1;
3169         return 0;
3170 }
3171
3172 static int drop_adjacent_extents(struct btrfs_trans_handle *trans,
3173                                  struct btrfs_root *root, struct inode *inode,
3174                                  struct extent_map *em,
3175                                  struct btrfs_path *path)
3176 {
3177         struct btrfs_file_extent_item *fi;
3178         struct extent_buffer *leaf;
3179         struct btrfs_key key, new_key;
3180         struct btrfs_map_token token;
3181         u64 extent_end;
3182         u64 extent_offset = 0;
3183         int extent_type;
3184         int del_slot = 0;
3185         int del_nr = 0;
3186         int ret = 0;
3187
3188         while (1) {
3189                 btrfs_init_map_token(&token);
3190                 leaf = path->nodes[0];
3191                 path->slots[0]++;
3192                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
3193                         if (del_nr) {
3194                                 ret = btrfs_del_items(trans, root, path,
3195                                                       del_slot, del_nr);
3196                                 if (ret)
3197                                         return ret;
3198                                 del_nr = 0;
3199                         }
3200
3201                         ret = btrfs_next_leaf_write(trans, root, path, 1);
3202                         if (ret < 0)
3203                                 return ret;
3204                         if (ret > 0)
3205                                 return 0;
3206                         leaf = path->nodes[0];
3207                 }
3208
3209                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3210                 if (key.objectid != btrfs_ino(inode) ||
3211                     key.type != BTRFS_EXTENT_DATA_KEY ||
3212                     key.offset >= em->start + em->len)
3213                         break;
3214
3215                 fi = btrfs_item_ptr(leaf, path->slots[0],
3216                                     struct btrfs_file_extent_item);
3217                 extent_type = btrfs_token_file_extent_type(leaf, fi, &token);
3218                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
3219                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
3220                         extent_offset = btrfs_token_file_extent_offset(leaf,
3221                                                                 fi, &token);
3222                         extent_end = key.offset +
3223                                 btrfs_token_file_extent_num_bytes(leaf, fi,
3224                                                                   &token);
3225                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
3226                         extent_end = key.offset +
3227                                 btrfs_file_extent_inline_len(leaf, fi);
3228                 } else {
3229                         BUG();
3230                 }
3231
3232                 if (extent_end <= em->len + em->start) {
3233                         if (!del_nr) {
3234                                 del_slot = path->slots[0];
3235                         }
3236                         del_nr++;
3237                         continue;
3238                 }
3239
3240                 /*
3241                  * Ok so we'll ignore previous items if we log a new extent,
3242                  * which can lead to overlapping extents, so if we have an
3243                  * existing extent we want to adjust we _have_ to check the next
3244                  * guy to make sure we even need this extent anymore, this keeps
3245                  * us from panicing in set_item_key_safe.
3246                  */
3247                 if (path->slots[0] < btrfs_header_nritems(leaf) - 1) {
3248                         struct btrfs_key tmp_key;
3249
3250                         btrfs_item_key_to_cpu(leaf, &tmp_key,
3251                                               path->slots[0] + 1);
3252                         if (tmp_key.objectid == btrfs_ino(inode) &&
3253                             tmp_key.type == BTRFS_EXTENT_DATA_KEY &&
3254                             tmp_key.offset <= em->start + em->len) {
3255                                 if (!del_nr)
3256                                         del_slot = path->slots[0];
3257                                 del_nr++;
3258                                 continue;
3259                         }
3260                 }
3261
3262                 BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
3263                 memcpy(&new_key, &key, sizeof(new_key));
3264                 new_key.offset = em->start + em->len;
3265                 btrfs_set_item_key_safe(trans, root, path, &new_key);
3266                 extent_offset += em->start + em->len - key.offset;
3267                 btrfs_set_token_file_extent_offset(leaf, fi, extent_offset,
3268                                                    &token);
3269                 btrfs_set_token_file_extent_num_bytes(leaf, fi, extent_end -
3270                                                       (em->start + em->len),
3271                                                       &token);
3272                 btrfs_mark_buffer_dirty(leaf);
3273         }
3274
3275         if (del_nr)
3276                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
3277
3278         return ret;
3279 }
3280
3281 static int log_one_extent(struct btrfs_trans_handle *trans,
3282                           struct inode *inode, struct btrfs_root *root,
3283                           struct extent_map *em, struct btrfs_path *path)
3284 {
3285         struct btrfs_root *log = root->log_root;
3286         struct btrfs_file_extent_item *fi;
3287         struct extent_buffer *leaf;
3288         struct btrfs_ordered_extent *ordered;
3289         struct list_head ordered_sums;
3290         struct btrfs_map_token token;
3291         struct btrfs_key key;
3292         u64 mod_start = em->mod_start;
3293         u64 mod_len = em->mod_len;
3294         u64 csum_offset;
3295         u64 csum_len;
3296         u64 extent_offset = em->start - em->orig_start;
3297         u64 block_len;
3298         int ret;
3299         int index = log->log_transid % 2;
3300         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
3301
3302         INIT_LIST_HEAD(&ordered_sums);
3303         btrfs_init_map_token(&token);
3304         key.objectid = btrfs_ino(inode);
3305         key.type = BTRFS_EXTENT_DATA_KEY;
3306         key.offset = em->start;
3307         path->really_keep_locks = 1;
3308
3309         ret = btrfs_insert_empty_item(trans, log, path, &key, sizeof(*fi));
3310         if (ret && ret != -EEXIST) {
3311                 path->really_keep_locks = 0;
3312                 return ret;
3313         }
3314         leaf = path->nodes[0];
3315         fi = btrfs_item_ptr(leaf, path->slots[0],
3316                             struct btrfs_file_extent_item);
3317         btrfs_set_token_file_extent_generation(leaf, fi, em->generation,
3318                                                &token);
3319         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3320                 skip_csum = true;
3321                 btrfs_set_token_file_extent_type(leaf, fi,
3322                                                  BTRFS_FILE_EXTENT_PREALLOC,
3323                                                  &token);
3324         } else {
3325                 btrfs_set_token_file_extent_type(leaf, fi,
3326                                                  BTRFS_FILE_EXTENT_REG,
3327                                                  &token);
3328                 if (em->block_start == 0)
3329                         skip_csum = true;
3330         }
3331
3332         block_len = max(em->block_len, em->orig_block_len);
3333         if (em->compress_type != BTRFS_COMPRESS_NONE) {
3334                 btrfs_set_token_file_extent_disk_bytenr(leaf, fi,
3335                                                         em->block_start,
3336                                                         &token);
3337                 btrfs_set_token_file_extent_disk_num_bytes(leaf, fi, block_len,
3338                                                            &token);
3339         } else if (em->block_start < EXTENT_MAP_LAST_BYTE) {
3340                 btrfs_set_token_file_extent_disk_bytenr(leaf, fi,
3341                                                         em->block_start -
3342                                                         extent_offset, &token);
3343                 btrfs_set_token_file_extent_disk_num_bytes(leaf, fi, block_len,
3344                                                            &token);
3345         } else {
3346                 btrfs_set_token_file_extent_disk_bytenr(leaf, fi, 0, &token);
3347                 btrfs_set_token_file_extent_disk_num_bytes(leaf, fi, 0,
3348                                                            &token);
3349         }
3350
3351         btrfs_set_token_file_extent_offset(leaf, fi,
3352                                            em->start - em->orig_start,
3353                                            &token);
3354         btrfs_set_token_file_extent_num_bytes(leaf, fi, em->len, &token);
3355         btrfs_set_token_file_extent_ram_bytes(leaf, fi, em->len, &token);
3356         btrfs_set_token_file_extent_compression(leaf, fi, em->compress_type,
3357                                                 &token);
3358         btrfs_set_token_file_extent_encryption(leaf, fi, 0, &token);
3359         btrfs_set_token_file_extent_other_encoding(leaf, fi, 0, &token);
3360         btrfs_mark_buffer_dirty(leaf);
3361
3362         /*
3363          * Have to check the extent to the right of us to make sure it doesn't
3364          * fall in our current range.  We're ok if the previous extent is in our
3365          * range since the recovery stuff will run us in key order and thus just
3366          * drop the part we overwrote.
3367          */
3368         ret = drop_adjacent_extents(trans, log, inode, em, path);
3369         btrfs_release_path(path);
3370         path->really_keep_locks = 0;
3371         if (ret) {
3372                 return ret;
3373         }
3374
3375         if (skip_csum)
3376                 return 0;
3377
3378         if (em->compress_type) {
3379                 csum_offset = 0;
3380                 csum_len = block_len;
3381         }
3382
3383         /*
3384          * First check and see if our csums are on our outstanding ordered
3385          * extents.
3386          */
3387 again:
3388         spin_lock_irq(&log->log_extents_lock[index]);
3389         list_for_each_entry(ordered, &log->logged_list[index], log_list) {
3390                 struct btrfs_ordered_sum *sum;
3391
3392                 if (!mod_len)
3393                         break;
3394
3395                 if (ordered->inode != inode)
3396                         continue;
3397
3398                 if (ordered->file_offset + ordered->len <= mod_start ||
3399                     mod_start + mod_len <= ordered->file_offset)
3400                         continue;
3401
3402                 /*
3403                  * We are going to copy all the csums on this ordered extent, so
3404                  * go ahead and adjust mod_start and mod_len in case this
3405                  * ordered extent has already been logged.
3406                  */
3407                 if (ordered->file_offset > mod_start) {
3408                         if (ordered->file_offset + ordered->len >=
3409                             mod_start + mod_len)
3410                                 mod_len = ordered->file_offset - mod_start;
3411                         /*
3412                          * If we have this case
3413                          *
3414                          * |--------- logged extent ---------|
3415                          *       |----- ordered extent ----|
3416                          *
3417                          * Just don't mess with mod_start and mod_len, we'll
3418                          * just end up logging more csums than we need and it
3419                          * will be ok.
3420                          */
3421                 } else {
3422                         if (ordered->file_offset + ordered->len <
3423                             mod_start + mod_len) {
3424                                 mod_len = (mod_start + mod_len) -
3425                                         (ordered->file_offset + ordered->len);
3426                                 mod_start = ordered->file_offset +
3427                                         ordered->len;
3428                         } else {
3429                                 mod_len = 0;
3430                         }
3431                 }
3432
3433                 /*
3434                  * To keep us from looping for the above case of an ordered
3435                  * extent that falls inside of the logged extent.
3436                  */
3437                 if (test_and_set_bit(BTRFS_ORDERED_LOGGED_CSUM,
3438                                      &ordered->flags))
3439                         continue;
3440                 atomic_inc(&ordered->refs);
3441                 spin_unlock_irq(&log->log_extents_lock[index]);
3442                 /*
3443                  * we've dropped the lock, we must either break or
3444                  * start over after this.
3445                  */
3446
3447                 wait_event(ordered->wait, ordered->csum_bytes_left == 0);
3448
3449                 list_for_each_entry(sum, &ordered->list, list) {
3450                         ret = btrfs_csum_file_blocks(trans, log, sum);
3451                         if (ret) {
3452                                 btrfs_put_ordered_extent(ordered);
3453                                 goto unlocked;
3454                         }
3455                 }
3456                 btrfs_put_ordered_extent(ordered);
3457                 goto again;
3458
3459         }
3460         spin_unlock_irq(&log->log_extents_lock[index]);
3461 unlocked:
3462
3463         if (!mod_len || ret)
3464                 return ret;
3465
3466         csum_offset = mod_start - em->start;
3467         csum_len = mod_len;
3468
3469         /* block start is already adjusted for the file extent offset. */
3470         ret = btrfs_lookup_csums_range(log->fs_info->csum_root,
3471                                        em->block_start + csum_offset,
3472                                        em->block_start + csum_offset +
3473                                        csum_len - 1, &ordered_sums, 0);
3474         if (ret)
3475                 return ret;
3476
3477         while (!list_empty(&ordered_sums)) {
3478                 struct btrfs_ordered_sum *sums = list_entry(ordered_sums.next,
3479                                                    struct btrfs_ordered_sum,
3480                                                    list);
3481                 if (!ret)
3482                         ret = btrfs_csum_file_blocks(trans, log, sums);
3483                 list_del(&sums->list);
3484                 kfree(sums);
3485         }
3486
3487         return ret;
3488 }
3489
3490 static int btrfs_log_changed_extents(struct btrfs_trans_handle *trans,
3491                                      struct btrfs_root *root,
3492                                      struct inode *inode,
3493                                      struct btrfs_path *path)
3494 {
3495         struct extent_map *em, *n;
3496         struct list_head extents;
3497         struct extent_map_tree *tree = &BTRFS_I(inode)->extent_tree;
3498         u64 test_gen;
3499         int ret = 0;
3500         int num = 0;
3501
3502         INIT_LIST_HEAD(&extents);
3503
3504         write_lock(&tree->lock);
3505         test_gen = root->fs_info->last_trans_committed;
3506
3507         list_for_each_entry_safe(em, n, &tree->modified_extents, list) {
3508                 list_del_init(&em->list);
3509
3510                 /*
3511                  * Just an arbitrary number, this can be really CPU intensive
3512                  * once we start getting a lot of extents, and really once we
3513                  * have a bunch of extents we just want to commit since it will
3514                  * be faster.
3515                  */
3516                 if (++num > 32768) {
3517                         list_del_init(&tree->modified_extents);
3518                         ret = -EFBIG;
3519                         goto process;
3520                 }
3521
3522                 if (em->generation <= test_gen)
3523                         continue;
3524                 /* Need a ref to keep it from getting evicted from cache */
3525                 atomic_inc(&em->refs);
3526                 set_bit(EXTENT_FLAG_LOGGING, &em->flags);
3527                 list_add_tail(&em->list, &extents);
3528                 num++;
3529         }
3530
3531         list_sort(NULL, &extents, extent_cmp);
3532
3533 process:
3534         while (!list_empty(&extents)) {
3535                 em = list_entry(extents.next, struct extent_map, list);
3536
3537                 list_del_init(&em->list);
3538
3539                 /*
3540                  * If we had an error we just need to delete everybody from our
3541                  * private list.
3542                  */
3543                 if (ret) {
3544                         clear_em_logging(tree, em);
3545                         free_extent_map(em);
3546                         continue;
3547                 }
3548
3549                 write_unlock(&tree->lock);
3550
3551                 ret = log_one_extent(trans, inode, root, em, path);
3552                 write_lock(&tree->lock);
3553                 clear_em_logging(tree, em);
3554                 free_extent_map(em);
3555         }
3556         WARN_ON(!list_empty(&extents));
3557         write_unlock(&tree->lock);
3558
3559         btrfs_release_path(path);
3560         return ret;
3561 }
3562
3563 /* log a single inode in the tree log.
3564  * At least one parent directory for this inode must exist in the tree
3565  * or be logged already.
3566  *
3567  * Any items from this inode changed by the current transaction are copied
3568  * to the log tree.  An extra reference is taken on any extents in this
3569  * file, allowing us to avoid a whole pile of corner cases around logging
3570  * blocks that have been removed from the tree.
3571  *
3572  * See LOG_INODE_ALL and related defines for a description of what inode_only
3573  * does.
3574  *
3575  * This handles both files and directories.
3576  */
3577 static int btrfs_log_inode(struct btrfs_trans_handle *trans,
3578                              struct btrfs_root *root, struct inode *inode,
3579                              int inode_only)
3580 {
3581         struct btrfs_path *path;
3582         struct btrfs_path *dst_path;
3583         struct btrfs_key min_key;
3584         struct btrfs_key max_key;
3585         struct btrfs_root *log = root->log_root;
3586         struct extent_buffer *src = NULL;
3587         int err = 0;
3588         int ret;
3589         int nritems;
3590         int ins_start_slot = 0;
3591         int ins_nr;
3592         bool fast_search = false;
3593         u64 ino = btrfs_ino(inode);
3594
3595         log = root->log_root;
3596
3597         path = btrfs_alloc_path();
3598         if (!path)
3599                 return -ENOMEM;
3600         dst_path = btrfs_alloc_path();
3601         if (!dst_path) {
3602                 btrfs_free_path(path);
3603                 return -ENOMEM;
3604         }
3605
3606         min_key.objectid = ino;
3607         min_key.type = BTRFS_INODE_ITEM_KEY;
3608         min_key.offset = 0;
3609
3610         max_key.objectid = ino;
3611
3612
3613         /* today the code can only do partial logging of directories */
3614         if (S_ISDIR(inode->i_mode) ||
3615             (!test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3616                        &BTRFS_I(inode)->runtime_flags) &&
3617              inode_only == LOG_INODE_EXISTS))
3618                 max_key.type = BTRFS_XATTR_ITEM_KEY;
3619         else
3620                 max_key.type = (u8)-1;
3621         max_key.offset = (u64)-1;
3622
3623         /* Only run delayed items if we are a dir or a new file */
3624         if (S_ISDIR(inode->i_mode) ||
3625             BTRFS_I(inode)->generation > root->fs_info->last_trans_committed) {
3626                 ret = btrfs_commit_inode_delayed_items(trans, inode);
3627                 if (ret) {
3628                         btrfs_free_path(path);
3629                         btrfs_free_path(dst_path);
3630                         return ret;
3631                 }
3632         }
3633
3634         mutex_lock(&BTRFS_I(inode)->log_mutex);
3635
3636         btrfs_get_logged_extents(log, inode);
3637
3638         /*
3639          * a brute force approach to making sure we get the most uptodate
3640          * copies of everything.
3641          */
3642         if (S_ISDIR(inode->i_mode)) {
3643                 int max_key_type = BTRFS_DIR_LOG_INDEX_KEY;
3644
3645                 if (inode_only == LOG_INODE_EXISTS)
3646                         max_key_type = BTRFS_XATTR_ITEM_KEY;
3647                 ret = drop_objectid_items(trans, log, path, ino, max_key_type);
3648         } else {
3649                 if (test_and_clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3650                                        &BTRFS_I(inode)->runtime_flags)) {
3651                         clear_bit(BTRFS_INODE_COPY_EVERYTHING,
3652                                   &BTRFS_I(inode)->runtime_flags);
3653                         ret = btrfs_truncate_inode_items(trans, log,
3654                                                          inode, 0, 0);
3655                 } else if (test_and_clear_bit(BTRFS_INODE_COPY_EVERYTHING,
3656                                               &BTRFS_I(inode)->runtime_flags)) {
3657                         if (inode_only == LOG_INODE_ALL)
3658                                 fast_search = true;
3659                         max_key.type = BTRFS_XATTR_ITEM_KEY;
3660                         ret = drop_objectid_items(trans, log, path, ino,
3661                                                   max_key.type);
3662                 } else {
3663                         if (inode_only == LOG_INODE_ALL)
3664                                 fast_search = true;
3665                         ret = log_inode_item(trans, log, dst_path, inode);
3666                         if (ret) {
3667                                 err = ret;
3668                                 goto out_unlock;
3669                         }
3670                         goto log_extents;
3671                 }
3672
3673         }
3674         if (ret) {
3675                 err = ret;
3676                 goto out_unlock;
3677         }
3678         path->keep_locks = 1;
3679
3680         while (1) {
3681                 ins_nr = 0;
3682                 ret = btrfs_search_forward(root, &min_key, &max_key,
3683                                            path, trans->transid);
3684                 if (ret != 0)
3685                         break;
3686 again:
3687                 /* note, ins_nr might be > 0 here, cleanup outside the loop */
3688                 if (min_key.objectid != ino)
3689                         break;
3690                 if (min_key.type > max_key.type)
3691                         break;
3692
3693                 src = path->nodes[0];
3694                 if (ins_nr && ins_start_slot + ins_nr == path->slots[0]) {
3695                         ins_nr++;
3696                         goto next_slot;
3697                 } else if (!ins_nr) {
3698                         ins_start_slot = path->slots[0];
3699                         ins_nr = 1;
3700                         goto next_slot;
3701                 }
3702
3703                 ret = copy_items(trans, inode, dst_path, src, ins_start_slot,
3704                                  ins_nr, inode_only);
3705                 if (ret) {
3706                         err = ret;
3707                         goto out_unlock;
3708                 }
3709                 ins_nr = 1;
3710                 ins_start_slot = path->slots[0];
3711 next_slot:
3712
3713                 nritems = btrfs_header_nritems(path->nodes[0]);
3714                 path->slots[0]++;
3715                 if (path->slots[0] < nritems) {
3716                         btrfs_item_key_to_cpu(path->nodes[0], &min_key,
3717                                               path->slots[0]);
3718                         goto again;
3719                 }
3720                 if (ins_nr) {
3721                         ret = copy_items(trans, inode, dst_path, src,
3722                                          ins_start_slot,
3723                                          ins_nr, inode_only);
3724                         if (ret) {
3725                                 err = ret;
3726                                 goto out_unlock;
3727                         }
3728                         ins_nr = 0;
3729                 }
3730                 btrfs_release_path(path);
3731
3732                 if (min_key.offset < (u64)-1)
3733                         min_key.offset++;
3734                 else if (min_key.type < (u8)-1)
3735                         min_key.type++;
3736                 else if (min_key.objectid < (u64)-1)
3737                         min_key.objectid++;
3738                 else
3739                         break;
3740         }
3741         if (ins_nr) {
3742                 ret = copy_items(trans, inode, dst_path, src, ins_start_slot,
3743                                  ins_nr, inode_only);
3744                 if (ret) {
3745                         err = ret;
3746                         goto out_unlock;
3747                 }
3748                 ins_nr = 0;
3749         }
3750
3751 log_extents:
3752         if (fast_search) {
3753                 btrfs_release_path(dst_path);
3754                 ret = btrfs_log_changed_extents(trans, root, inode, dst_path);
3755                 if (ret) {
3756                         err = ret;
3757                         goto out_unlock;
3758                 }
3759         } else {
3760                 struct extent_map_tree *tree = &BTRFS_I(inode)->extent_tree;
3761                 struct extent_map *em, *n;
3762
3763                 write_lock(&tree->lock);
3764                 list_for_each_entry_safe(em, n, &tree->modified_extents, list)
3765                         list_del_init(&em->list);
3766                 write_unlock(&tree->lock);
3767         }
3768
3769         if (inode_only == LOG_INODE_ALL && S_ISDIR(inode->i_mode)) {
3770                 btrfs_release_path(path);
3771                 btrfs_release_path(dst_path);
3772                 ret = log_directory_changes(trans, root, inode, path, dst_path);
3773                 if (ret) {
3774                         err = ret;
3775                         goto out_unlock;
3776                 }
3777         }
3778         BTRFS_I(inode)->logged_trans = trans->transid;
3779         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->last_sub_trans;
3780 out_unlock:
3781         if (err)
3782                 btrfs_free_logged_extents(log, log->log_transid);
3783         mutex_unlock(&BTRFS_I(inode)->log_mutex);
3784
3785         btrfs_free_path(path);
3786         btrfs_free_path(dst_path);
3787         return err;
3788 }
3789
3790 /*
3791  * follow the dentry parent pointers up the chain and see if any
3792  * of the directories in it require a full commit before they can
3793  * be logged.  Returns zero if nothing special needs to be done or 1 if
3794  * a full commit is required.
3795  */
3796 static noinline int check_parent_dirs_for_sync(struct btrfs_trans_handle *trans,
3797                                                struct inode *inode,
3798                                                struct dentry *parent,
3799                                                struct super_block *sb,
3800                                                u64 last_committed)
3801 {
3802         int ret = 0;
3803         struct btrfs_root *root;
3804         struct dentry *old_parent = NULL;
3805
3806         /*
3807          * for regular files, if its inode is already on disk, we don't
3808          * have to worry about the parents at all.  This is because
3809          * we can use the last_unlink_trans field to record renames
3810          * and other fun in this file.
3811          */
3812         if (S_ISREG(inode->i_mode) &&
3813             BTRFS_I(inode)->generation <= last_committed &&
3814             BTRFS_I(inode)->last_unlink_trans <= last_committed)
3815                         goto out;
3816
3817         if (!S_ISDIR(inode->i_mode)) {
3818                 if (!parent || !parent->d_inode || sb != parent->d_inode->i_sb)
3819                         goto out;
3820                 inode = parent->d_inode;
3821         }
3822
3823         while (1) {
3824                 BTRFS_I(inode)->logged_trans = trans->transid;
3825                 smp_mb();
3826
3827                 if (BTRFS_I(inode)->last_unlink_trans > last_committed) {
3828                         root = BTRFS_I(inode)->root;
3829
3830                         /*
3831                          * make sure any commits to the log are forced
3832                          * to be full commits
3833                          */
3834                         root->fs_info->last_trans_log_full_commit =
3835                                 trans->transid;
3836                         ret = 1;
3837                         break;
3838                 }
3839
3840                 if (!parent || !parent->d_inode || sb != parent->d_inode->i_sb)
3841                         break;
3842
3843                 if (IS_ROOT(parent))
3844                         break;
3845
3846                 parent = dget_parent(parent);
3847                 dput(old_parent);
3848                 old_parent = parent;
3849                 inode = parent->d_inode;
3850
3851         }
3852         dput(old_parent);
3853 out:
3854         return ret;
3855 }
3856
3857 /*
3858  * helper function around btrfs_log_inode to make sure newly created
3859  * parent directories also end up in the log.  A minimal inode and backref
3860  * only logging is done of any parent directories that are older than
3861  * the last committed transaction
3862  */
3863 int btrfs_log_inode_parent(struct btrfs_trans_handle *trans,
3864                     struct btrfs_root *root, struct inode *inode,
3865                     struct dentry *parent, int exists_only)
3866 {
3867         int inode_only = exists_only ? LOG_INODE_EXISTS : LOG_INODE_ALL;
3868         struct super_block *sb;
3869         struct dentry *old_parent = NULL;
3870         int ret = 0;
3871         u64 last_committed = root->fs_info->last_trans_committed;
3872
3873         sb = inode->i_sb;
3874
3875         if (btrfs_test_opt(root, NOTREELOG)) {
3876                 ret = 1;
3877                 goto end_no_trans;
3878         }
3879
3880         if (root->fs_info->last_trans_log_full_commit >
3881             root->fs_info->last_trans_committed) {
3882                 ret = 1;
3883                 goto end_no_trans;
3884         }
3885
3886         if (root != BTRFS_I(inode)->root ||
3887             btrfs_root_refs(&root->root_item) == 0) {
3888                 ret = 1;
3889                 goto end_no_trans;
3890         }
3891
3892         ret = check_parent_dirs_for_sync(trans, inode, parent,
3893                                          sb, last_committed);
3894         if (ret)
3895                 goto end_no_trans;
3896
3897         if (btrfs_inode_in_log(inode, trans->transid)) {
3898                 ret = BTRFS_NO_LOG_SYNC;
3899                 goto end_no_trans;
3900         }
3901
3902         ret = start_log_trans(trans, root);
3903         if (ret)
3904                 goto end_trans;
3905
3906         ret = btrfs_log_inode(trans, root, inode, inode_only);
3907         if (ret)
3908                 goto end_trans;
3909
3910         /*
3911          * for regular files, if its inode is already on disk, we don't
3912          * have to worry about the parents at all.  This is because
3913          * we can use the last_unlink_trans field to record renames
3914          * and other fun in this file.
3915          */
3916         if (S_ISREG(inode->i_mode) &&
3917             BTRFS_I(inode)->generation <= last_committed &&
3918             BTRFS_I(inode)->last_unlink_trans <= last_committed) {
3919                 ret = 0;
3920                 goto end_trans;
3921         }
3922
3923         inode_only = LOG_INODE_EXISTS;
3924         while (1) {
3925                 if (!parent || !parent->d_inode || sb != parent->d_inode->i_sb)
3926                         break;
3927
3928                 inode = parent->d_inode;
3929                 if (root != BTRFS_I(inode)->root)
3930                         break;
3931
3932                 if (BTRFS_I(inode)->generation >
3933                     root->fs_info->last_trans_committed) {
3934                         ret = btrfs_log_inode(trans, root, inode, inode_only);
3935                         if (ret)
3936                                 goto end_trans;
3937                 }
3938                 if (IS_ROOT(parent))
3939                         break;
3940
3941                 parent = dget_parent(parent);
3942                 dput(old_parent);
3943                 old_parent = parent;
3944         }
3945         ret = 0;
3946 end_trans:
3947         dput(old_parent);
3948         if (ret < 0) {
3949                 root->fs_info->last_trans_log_full_commit = trans->transid;
3950                 ret = 1;
3951         }
3952         btrfs_end_log_trans(root);
3953 end_no_trans:
3954         return ret;
3955 }
3956
3957 /*
3958  * it is not safe to log dentry if the chunk root has added new
3959  * chunks.  This returns 0 if the dentry was logged, and 1 otherwise.
3960  * If this returns 1, you must commit the transaction to safely get your
3961  * data on disk.
3962  */
3963 int btrfs_log_dentry_safe(struct btrfs_trans_handle *trans,
3964                           struct btrfs_root *root, struct dentry *dentry)
3965 {
3966         struct dentry *parent = dget_parent(dentry);
3967         int ret;
3968
3969         ret = btrfs_log_inode_parent(trans, root, dentry->d_inode, parent, 0);
3970         dput(parent);
3971
3972         return ret;
3973 }
3974
3975 /*
3976  * should be called during mount to recover any replay any log trees
3977  * from the FS
3978  */
3979 int btrfs_recover_log_trees(struct btrfs_root *log_root_tree)
3980 {
3981         int ret;
3982         struct btrfs_path *path;
3983         struct btrfs_trans_handle *trans;
3984         struct btrfs_key key;
3985         struct btrfs_key found_key;
3986         struct btrfs_key tmp_key;
3987         struct btrfs_root *log;
3988         struct btrfs_fs_info *fs_info = log_root_tree->fs_info;
3989         struct walk_control wc = {
3990                 .process_func = process_one_buffer,
3991                 .stage = 0,
3992         };
3993
3994         path = btrfs_alloc_path();
3995         if (!path)
3996                 return -ENOMEM;
3997
3998         fs_info->log_root_recovering = 1;
3999
4000         trans = btrfs_start_transaction(fs_info->tree_root, 0);
4001         if (IS_ERR(trans)) {
4002                 ret = PTR_ERR(trans);
4003                 goto error;
4004         }
4005
4006         wc.trans = trans;
4007         wc.pin = 1;
4008
4009         ret = walk_log_tree(trans, log_root_tree, &wc);
4010         if (ret) {
4011                 btrfs_error(fs_info, ret, "Failed to pin buffers while "
4012                             "recovering log root tree.");
4013                 goto error;
4014         }
4015
4016 again:
4017         key.objectid = BTRFS_TREE_LOG_OBJECTID;
4018         key.offset = (u64)-1;
4019         btrfs_set_key_type(&key, BTRFS_ROOT_ITEM_KEY);
4020
4021         while (1) {
4022                 ret = btrfs_search_slot(NULL, log_root_tree, &key, path, 0, 0);
4023
4024                 if (ret < 0) {
4025                         btrfs_error(fs_info, ret,
4026                                     "Couldn't find tree log root.");
4027                         goto error;
4028                 }
4029                 if (ret > 0) {
4030                         if (path->slots[0] == 0)
4031                                 break;
4032                         path->slots[0]--;
4033                 }
4034                 btrfs_item_key_to_cpu(path->nodes[0], &found_key,
4035                                       path->slots[0]);
4036                 btrfs_release_path(path);
4037                 if (found_key.objectid != BTRFS_TREE_LOG_OBJECTID)
4038                         break;
4039
4040                 log = btrfs_read_fs_root_no_radix(log_root_tree,
4041                                                   &found_key);
4042                 if (IS_ERR(log)) {
4043                         ret = PTR_ERR(log);
4044                         btrfs_error(fs_info, ret,
4045                                     "Couldn't read tree log root.");
4046                         goto error;
4047                 }
4048
4049                 tmp_key.objectid = found_key.offset;
4050                 tmp_key.type = BTRFS_ROOT_ITEM_KEY;
4051                 tmp_key.offset = (u64)-1;
4052
4053                 wc.replay_dest = btrfs_read_fs_root_no_name(fs_info, &tmp_key);
4054                 if (IS_ERR(wc.replay_dest)) {
4055                         ret = PTR_ERR(wc.replay_dest);
4056                         btrfs_error(fs_info, ret, "Couldn't read target root "
4057                                     "for tree log recovery.");
4058                         goto error;
4059                 }
4060
4061                 wc.replay_dest->log_root = log;
4062                 btrfs_record_root_in_trans(trans, wc.replay_dest);
4063                 ret = walk_log_tree(trans, log, &wc);
4064                 BUG_ON(ret);
4065
4066                 if (wc.stage == LOG_WALK_REPLAY_ALL) {
4067                         ret = fixup_inode_link_counts(trans, wc.replay_dest,
4068                                                       path);
4069                         BUG_ON(ret);
4070                 }
4071
4072                 key.offset = found_key.offset - 1;
4073                 wc.replay_dest->log_root = NULL;
4074                 free_extent_buffer(log->node);
4075                 free_extent_buffer(log->commit_root);
4076                 kfree(log);
4077
4078                 if (found_key.offset == 0)
4079                         break;
4080         }
4081         btrfs_release_path(path);
4082
4083         /* step one is to pin it all, step two is to replay just inodes */
4084         if (wc.pin) {
4085                 wc.pin = 0;
4086                 wc.process_func = replay_one_buffer;
4087                 wc.stage = LOG_WALK_REPLAY_INODES;
4088                 goto again;
4089         }
4090         /* step three is to replay everything */
4091         if (wc.stage < LOG_WALK_REPLAY_ALL) {
4092                 wc.stage++;
4093                 goto again;
4094         }
4095
4096         btrfs_free_path(path);
4097
4098         free_extent_buffer(log_root_tree->node);
4099         log_root_tree->log_root = NULL;
4100         fs_info->log_root_recovering = 0;
4101
4102         /* step 4: commit the transaction, which also unpins the blocks */
4103         btrfs_commit_transaction(trans, fs_info->tree_root);
4104
4105         kfree(log_root_tree);
4106         return 0;
4107
4108 error:
4109         btrfs_free_path(path);
4110         return ret;
4111 }
4112
4113 /*
4114  * there are some corner cases where we want to force a full
4115  * commit instead of allowing a directory to be logged.
4116  *
4117  * They revolve around files there were unlinked from the directory, and
4118  * this function updates the parent directory so that a full commit is
4119  * properly done if it is fsync'd later after the unlinks are done.
4120  */
4121 void btrfs_record_unlink_dir(struct btrfs_trans_handle *trans,
4122                              struct inode *dir, struct inode *inode,
4123                              int for_rename)
4124 {
4125         /*
4126          * when we're logging a file, if it hasn't been renamed
4127          * or unlinked, and its inode is fully committed on disk,
4128          * we don't have to worry about walking up the directory chain
4129          * to log its parents.
4130          *
4131          * So, we use the last_unlink_trans field to put this transid
4132          * into the file.  When the file is logged we check it and
4133          * don't log the parents if the file is fully on disk.
4134          */
4135         if (S_ISREG(inode->i_mode))
4136                 BTRFS_I(inode)->last_unlink_trans = trans->transid;
4137
4138         /*
4139          * if this directory was already logged any new
4140          * names for this file/dir will get recorded
4141          */
4142         smp_mb();
4143         if (BTRFS_I(dir)->logged_trans == trans->transid)
4144                 return;
4145
4146         /*
4147          * if the inode we're about to unlink was logged,
4148          * the log will be properly updated for any new names
4149          */
4150         if (BTRFS_I(inode)->logged_trans == trans->transid)
4151                 return;
4152
4153         /*
4154          * when renaming files across directories, if the directory
4155          * there we're unlinking from gets fsync'd later on, there's
4156          * no way to find the destination directory later and fsync it
4157          * properly.  So, we have to be conservative and force commits
4158          * so the new name gets discovered.
4159          */
4160         if (for_rename)
4161                 goto record;
4162
4163         /* we can safely do the unlink without any special recording */
4164         return;
4165
4166 record:
4167         BTRFS_I(dir)->last_unlink_trans = trans->transid;
4168 }
4169
4170 /*
4171  * Call this after adding a new name for a file and it will properly
4172  * update the log to reflect the new name.
4173  *
4174  * It will return zero if all goes well, and it will return 1 if a
4175  * full transaction commit is required.
4176  */
4177 int btrfs_log_new_name(struct btrfs_trans_handle *trans,
4178                         struct inode *inode, struct inode *old_dir,
4179                         struct dentry *parent)
4180 {
4181         struct btrfs_root * root = BTRFS_I(inode)->root;
4182
4183         /*
4184          * this will force the logging code to walk the dentry chain
4185          * up for the file
4186          */
4187         if (S_ISREG(inode->i_mode))
4188                 BTRFS_I(inode)->last_unlink_trans = trans->transid;
4189
4190         /*
4191          * if this inode hasn't been logged and directory we're renaming it
4192          * from hasn't been logged, we don't need to log it
4193          */
4194         if (BTRFS_I(inode)->logged_trans <=
4195             root->fs_info->last_trans_committed &&
4196             (!old_dir || BTRFS_I(old_dir)->logged_trans <=
4197                     root->fs_info->last_trans_committed))
4198                 return 0;
4199
4200         return btrfs_log_inode_parent(trans, root, inode, parent, 1);
4201 }
4202